Uploaded by Juan Josafat Zetera

Machine Learning with Python Cookbook Practical Solutions from Preprocessing to Deep Learning by Chris Albon (z-lib.org)

advertisement
Python
Machine
Learning
Cookbook
PRACTICAL SOLUTIONS FROM PREPROCESSING TO DEEP LEARNING
Chris Albon
Machine Learning with
Python Cookbook
Practical Solutions from Preprocessing
to Deep Learning
Chris Albon
Machine Learning with Python Cookbook
by Chris Albon
Copyright © 2018 Chris Albon. All rights reserved.
Printed in the United States of America.
Published by O’Reilly Media, Inc., 1005 Gravenstein Highway North, Sebastopol, CA 95472.
O’Reilly books may be purchased for educational, business, or sales promotional use. Online editions are
also available for most titles (http://oreilly.com/safari). For more information, contact our corporate/insti‐
tutional sales department: 800-998-9938 or corporate@oreilly.com.
Editors: Rachel Roumeliotis and Jeff Bleiel
Production Editor: Melanie Yarbrough
Copyeditor: Kim Cofer
Proofreader: Rachel Monaghan
April 2018:
Indexer: Wendy Catalano
Interior Designer: David Futato
Cover Designer: Karen Montgomery
Illustrator: Rebecca Demarest
First Edition
Revision History for the First Edition
2018-03-09:
First Release
See http://oreilly.com/catalog/errata.csp?isbn=9781491989388 for release details.
The O’Reilly logo is a registered trademark of O’Reilly Media, Inc. Machine Learning with Python Cook‐
book, the cover image, and related trade dress are trademarks of O’Reilly Media, Inc.
While the publisher and the author have used good faith efforts to ensure that the information and
instructions contained in this work are accurate, the publisher and the author disclaim all responsibility
for errors or omissions, including without limitation responsibility for damages resulting from the use of
or reliance on this work. Use of the information and instructions contained in this work is at your own
risk. If any code samples or other technology this work contains or describes is subject to open source
licenses or the intellectual property rights of others, it is your responsibility to ensure that your use
thereof complies with such licenses and/or rights.
978-1-491-98938-8
[LSI]
Table of Contents
Preface. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xi
1. Vectors, Matrices, and Arrays. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.0 Introduction
1.1 Creating a Vector
1.2 Creating a Matrix
1.3 Creating a Sparse Matrix
1.4 Selecting Elements
1.5 Describing a Matrix
1.6 Applying Operations to Elements
1.7 Finding the Maximum and Minimum Values
1.8 Calculating the Average, Variance, and Standard Deviation
1.9 Reshaping Arrays
1.10 Transposing a Vector or Matrix
1.11 Flattening a Matrix
1.12 Finding the Rank of a Matrix
1.13 Calculating the Determinant
1.14 Getting the Diagonal of a Matrix
1.15 Calculating the Trace of a Matrix
1.16 Finding Eigenvalues and Eigenvectors
1.17 Calculating Dot Products
1.18 Adding and Subtracting Matrices
1.19 Multiplying Matrices
1.20 Inverting a Matrix
1.21 Generating Random Values
1
1
2
3
4
6
6
7
8
9
10
11
12
12
13
14
15
16
17
18
19
20
2. Loading Data. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
2.0 Introduction
23
iii
2.1 Loading a Sample Dataset
2.2 Creating a Simulated Dataset
2.3 Loading a CSV File
2.4 Loading an Excel File
2.5 Loading a JSON File
2.6 Querying a SQL Database
23
24
27
28
29
30
3. Data Wrangling. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
3.0 Introduction
3.1 Creating a Data Frame
3.2 Describing the Data
3.3 Navigating DataFrames
3.4 Selecting Rows Based on Conditionals
3.5 Replacing Values
3.6 Renaming Columns
3.7 Finding the Minimum, Maximum, Sum, Average, and Count
3.8 Finding Unique Values
3.9 Handling Missing Values
3.10 Deleting a Column
3.11 Deleting a Row
3.12 Dropping Duplicate Rows
3.13 Grouping Rows by Values
3.14 Grouping Rows by Time
3.15 Looping Over a Column
3.16 Applying a Function Over All Elements in a Column
3.17 Applying a Function to Groups
3.18 Concatenating DataFrames
3.19 Merging DataFrames
33
34
35
37
38
39
41
42
43
44
46
47
48
50
51
53
54
55
55
57
4. Handling Numerical Data. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
4.0 Introduction
4.1 Rescaling a Feature
4.2 Standardizing a Feature
4.3 Normalizing Observations
4.4 Generating Polynomial and Interaction Features
4.5 Transforming Features
4.6 Detecting Outliers
4.7 Handling Outliers
4.8 Discretizating Features
4.9 Grouping Observations Using Clustering
4.10 Deleting Observations with Missing Values
4.11 Imputing Missing Values
iv
|
Table of Contents
61
61
63
64
66
68
69
71
73
74
76
78
5. Handling Categorical Data. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
5.0 Introduction
5.1 Encoding Nominal Categorical Features
5.2 Encoding Ordinal Categorical Features
5.3 Encoding Dictionaries of Features
5.4 Imputing Missing Class Values
5.5 Handling Imbalanced Classes
81
82
84
86
88
90
6. Handling Text. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
6.0 Introduction
6.1 Cleaning Text
6.2 Parsing and Cleaning HTML
6.3 Removing Punctuation
6.4 Tokenizing Text
6.5 Removing Stop Words
6.6 Stemming Words
6.7 Tagging Parts of Speech
6.8 Encoding Text as a Bag of Words
6.9 Weighting Word Importance
95
95
97
98
98
99
100
101
104
106
7. Handling Dates and Times. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
7.0 Introduction
7.1 Converting Strings to Dates
7.2 Handling Time Zones
7.3 Selecting Dates and Times
7.4 Breaking Up Date Data into Multiple Features
7.5 Calculating the Difference Between Dates
7.6 Encoding Days of the Week
7.7 Creating a Lagged Feature
7.8 Using Rolling Time Windows
7.9 Handling Missing Data in Time Series
109
109
111
112
113
114
115
116
117
118
8. Handling Images. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
8.0 Introduction
8.1 Loading Images
8.2 Saving Images
8.3 Resizing Images
8.4 Cropping Images
8.5 Blurring Images
8.6 Sharpening Images
8.7 Enhancing Contrast
8.8 Isolating Colors
121
122
124
125
126
128
131
133
135
Table of Contents
|
v
8.9 Binarizing Images
8.10 Removing Backgrounds
8.11 Detecting Edges
8.12 Detecting Corners
8.13 Creating Features for Machine Learning
8.14 Encoding Mean Color as a Feature
8.15 Encoding Color Histograms as Features
137
140
144
146
150
152
153
9. Dimensionality Reduction Using Feature Extraction. . . . . . . . . . . . . . . . . . . . . . . . . . . . 157
9.0 Introduction
9.1 Reducing Features Using Principal Components
9.2 Reducing Features When Data Is Linearly Inseparable
9.3 Reducing Features by Maximizing Class Separability
9.4 Reducing Features Using Matrix Factorization
9.5 Reducing Features on Sparse Data
157
158
160
162
165
166
10. Dimensionality Reduction Using Feature Selection. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 169
10.0 Introduction
10.1 Thresholding Numerical Feature Variance
10.2 Thresholding Binary Feature Variance
10.3 Handling Highly Correlated Features
10.4 Removing Irrelevant Features for Classification
10.5 Recursively Eliminating Features
169
170
171
172
174
176
11. Model Evaluation. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 179
11.0 Introduction
11.1 Cross-Validating Models
11.2 Creating a Baseline Regression Model
11.3 Creating a Baseline Classification Model
11.4 Evaluating Binary Classifier Predictions
11.5 Evaluating Binary Classifier Thresholds
11.6 Evaluating Multiclass Classifier Predictions
11.7 Visualizing a Classifier’s Performance
11.8 Evaluating Regression Models
11.9 Evaluating Clustering Models
11.10 Creating a Custom Evaluation Metric
11.11 Visualizing the Effect of Training Set Size
11.12 Creating a Text Report of Evaluation Metrics
11.13 Visualizing the Effect of Hyperparameter Values
179
179
183
184
186
189
192
194
196
198
199
201
203
205
12. Model Selection. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 209
12.0 Introduction
vi
|
Table of Contents
209
12.1 Selecting Best Models Using Exhaustive Search
12.2 Selecting Best Models Using Randomized Search
12.3 Selecting Best Models from Multiple Learning Algorithms
12.4 Selecting Best Models When Preprocessing
12.5 Speeding Up Model Selection with Parallelization
12.6 Speeding Up Model Selection Using Algorithm-Specific Methods
12.7 Evaluating Performance After Model Selection
210
212
214
215
217
219
220
13. Linear Regression. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 223
13.0 Introduction
13.1 Fitting a Line
13.2 Handling Interactive Effects
13.3 Fitting a Nonlinear Relationship
13.4 Reducing Variance with Regularization
13.5 Reducing Features with Lasso Regression
223
223
225
227
229
231
14. Trees and Forests. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 233
14.0 Introduction
14.1 Training a Decision Tree Classifier
14.2 Training a Decision Tree Regressor
14.3 Visualizing a Decision Tree Model
14.4 Training a Random Forest Classifier
14.5 Training a Random Forest Regressor
14.6 Identifying Important Features in Random Forests
14.7 Selecting Important Features in Random Forests
14.8 Handling Imbalanced Classes
14.9 Controlling Tree Size
14.10 Improving Performance Through Boosting
14.11 Evaluating Random Forests with Out-of-Bag Errors
233
233
235
236
238
240
241
243
245
246
247
249
15. K-Nearest Neighbors. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 251
15.0 Introduction
15.1 Finding an Observation’s Nearest Neighbors
15.2 Creating a K-Nearest Neighbor Classifier
15.3 Identifying the Best Neighborhood Size
15.4 Creating a Radius-Based Nearest Neighbor Classifier
251
251
254
256
257
16. Logistic Regression. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 259
16.0 Introduction
16.1 Training a Binary Classifier
16.2 Training a Multiclass Classifier
16.3 Reducing Variance Through Regularization
259
259
261
262
Table of Contents
|
vii
16.4 Training a Classifier on Very Large Data
16.5 Handling Imbalanced Classes
263
264
17. Support Vector Machines. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 267
17.0 Introduction
17.1 Training a Linear Classifier
17.2 Handling Linearly Inseparable Classes Using Kernels
17.3 Creating Predicted Probabilities
17.4 Identifying Support Vectors
17.5 Handling Imbalanced Classes
267
267
270
274
276
277
18. Naive Bayes. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 279
18.0 Introduction
18.1 Training a Classifier for Continuous Features
18.2 Training a Classifier for Discrete and Count Features
18.3 Training a Naive Bayes Classifier for Binary Features
18.4 Calibrating Predicted Probabilities
279
280
282
283
284
19. Clustering. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 287
19.0 Introduction
19.1 Clustering Using K-Means
19.2 Speeding Up K-Means Clustering
19.3 Clustering Using Meanshift
19.4 Clustering Using DBSCAN
19.5 Clustering Using Hierarchical Merging
287
287
290
291
292
294
20. Neural Networks. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 297
20.0 Introduction
20.1 Preprocessing Data for Neural Networks
20.2 Designing a Neural Network
20.3 Training a Binary Classifier
20.4 Training a Multiclass Classifier
20.5 Training a Regressor
20.6 Making Predictions
20.7 Visualize Training History
20.8 Reducing Overfitting with Weight Regularization
20.9 Reducing Overfitting with Early Stopping
20.10 Reducing Overfitting with Dropout
20.11 Saving Model Training Progress
20.12 k-Fold Cross-Validating Neural Networks
20.13 Tuning Neural Networks
20.14 Visualizing Neural Networks
viii
|
Table of Contents
297
298
300
303
305
307
309
310
313
315
317
319
321
322
325
20.15 Classifying Images
20.16 Improving Performance with Image Augmentation
20.17 Classifying Text
327
331
333
21. Saving and Loading Trained Models. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 337
21.0 Introduction
21.1 Saving and Loading a scikit-learn Model
21.2 Saving and Loading a Keras Model
337
337
339
Index. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 341
Table of Contents
|
ix
Preface
Over the last few years machine learning has become embedded in a wide variety of
day-to-day business, nonprofit, and government operations. As the popularity of
machine learning increased, a cottage industry of high-quality literature that taught
applied machine learning to practitioners developed. This literature has been highly
successful in training an entire generation of data scientists and machine learning
engineers. This literature also approached the topic of machine learning from the
perspective of providing a learning resource to teach an individual what machine
learning is and how it works. However, while fruitful, this approach left out a differ‐
ent perspective on the topic: the nuts and bolts of doing machine learning day to day.
That is the motivation of this book—not as a tome of machine learning knowledge
for the student but as a wrench for the professional, to sit with dog-eared pages on
desks ready to solve the practical day-to-day problems of a machine learning practi‐
tioner.
More specifically, the book takes a task-based approach to machine learning, with
almost 200 self-contained solutions (you can copy and paste the code and it’ll run)
for the most common tasks a data scientist or machine learning engineer building a
model will run into.
The ultimate goal is for the book to be a reference for people building real machine
learning systems. For example, imagine a reader has a JSON file containing 1,000 cat‐
egorical and numerical features with missing data and categorical target vectors with
imbalanced classes, and wants an interpretable model. The motivation for this book is
to provide recipes to help the reader learn processes such as:
• 2.5 Loading a JSON file
• 4.2 Standardizing a Feature
• 5.3 Encoding Dictionaries of Features
• 5.4 Imputing Missing Class Values
xi
• 9.1 Reducing Features Using Principal Components
• 12.2 Selecting Best Models Using Randomized Search
• 14.4 Training a Random Forest Classifier
• 14.7 Selecting Random Features in Random Forests
The goal is for the reader to be able to:
1. Copy/paste the code and gain confidence that it actually works with the included
toy dataset.
2. Read the discussion to gain an understanding of the theory behind the technique
the code is executing and learn which parameters are important to consider.
3. Insert/combine/adapt the code from the recipes to construct the actual applica‐
tion.
Who This Book Is For
This book is not an introduction to machine learning. If you are not comfortable with
the basic concepts of machine learning or have never spent time learning machine
learning, do not buy this book. Instead, this book is for the machine learning practi‐
tioner who, while comfortable with the theory and concepts of machine learning,
would benefit from a quick reference containing code to solve challenges he runs into
working on machine learning on an everyday basis.
This book assumes the reader is comfortable with the Python programming language
and package management.
Who This Book Is Not For
As stated previously, this book is not an introduction to machine learning. This book
should not be your first. If you are unfamiliar with concepts like cross-validation,
random forest, and gradient descent, you will likely not benefit from this book as
much as one of the many high-quality texts specifically designed to introduce you to
the topic. I recommend reading one of those books and then coming back to this
book to learn working, practical solutions for machine learning.
Terminology Used in This Book
Machine learning draws upon techniques from a wide range of fields, including com‐
puter science, statistics, and mathematics. For this reason, there is significant varia‐
tion in the terminology used in the discussions of machine learning:
xii
|
Preface
Observation
A single unit in our level of observation—for example, a person, a sale, or a
record.
Learning algorithms
An algorithm used to learn the best parameters of a model—for example, linear
regression, naive Bayes, or decision trees.
Models
An output of a learning algorithm’s training. Learning algorithms train models,
which we then use to make predictions.
Parameters
The weights or coefficients of a model learned through training.
Hyperparameters
The settings of a learning algorithm that need to be set before training.
Performance
A metric used to evaluate a model.
Loss
A metric to maximize or minimize through training.
Train
Applying a learning algorithm to data using numerical approaches like gradient
descent.
Fit
Applying a learning algorithm to data using analytical approaches.
Data
A collection of observations.
Acknowledgments
This book would not have been possible without the kind help of a number of friends
and strangers. Listing everyone who lent a hand to this project would be impossible,
but I wanted to at least mention: Angela Bassa, Teresa Borcuch, Justin Bozonier,
Andre deBruin, Numa Dhamani, Dan Friedman, Joel Grus, Sarah Guido, Bill Kam‐
bouroglou, Mat Kelcey, Lizzie Kumar, Hilary Parker, Niti Paudyal, Sebastian Raschka,
and Shreya Shankar.
I owe them all a beer or five.
Preface
|
xiii
CHAPTER 1
Vectors, Matrices, and Arrays
1.0 Introduction
NumPy is the foundation of the Python machine learning stack. NumPy allows for
efficient operations on the data structures often used in machine learning: vectors,
matrices, and tensors. While NumPy is not the focus of this book, it will show up fre‐
quently throughout the following chapters. This chapter covers the most common
NumPy operations we are likely to run into while working on machine learning
workflows.
1.1 Creating a Vector
Problem
You need to create a vector.
Solution
Use NumPy to create a one-dimensional array:
# Load library
import numpy as np
# Create a vector as a row
vector_row = np.array([1, 2, 3])
# Create a vector as a column
vector_column = np.array([[1],
[2],
[3]])
1
Discussion
NumPy’s main data structure is the multidimensional array. To create a vector, we
simply create a one-dimensional array. Just like vectors, these arrays can be repre‐
sented horizontally (i.e., rows) or vertically (i.e., columns).
See Also
• Vectors, Math Is Fun
• Euclidean vector, Wikipedia
1.2 Creating a Matrix
Problem
You need to create a matrix.
Solution
Use NumPy to create a two-dimensional array:
# Load library
import numpy as np
# Create a matrix
matrix = np.array([[1, 2],
[1, 2],
[1, 2]])
Discussion
To create a matrix we can use a NumPy two-dimensional array. In our solution, the
matrix contains three rows and two columns (a column of 1s and a column of 2s).
NumPy actually has a dedicated matrix data structure:
matrix_object = np.mat([[1, 2],
[1, 2],
[1, 2]])
matrix([[1, 2],
[1, 2],
[1, 2]])
However, the matrix data structure is not recommended for two reasons. First, arrays
are the de facto standard data structure of NumPy. Second, the vast majority of
NumPy operations return arrays, not matrix objects.
2
|
Chapter 1: Vectors, Matrices, and Arrays
See Also
• Matrix, Wikipedia
• Matrix, Wolfram MathWorld
1.3 Creating a Sparse Matrix
Problem
Given data with very few nonzero values, you want to efficiently represent it.
Solution
Create a sparse matrix:
# Load libraries
import numpy as np
from scipy import sparse
# Create a matrix
matrix = np.array([[0, 0],
[0, 1],
[3, 0]])
# Create compressed sparse row (CSR) matrix
matrix_sparse = sparse.csr_matrix(matrix)
Discussion
A frequent situation in machine learning is having a huge amount of data; however,
most of the elements in the data are zeros. For example, imagine a matrix where the
columns are every movie on Netflix, the rows are every Netflix user, and the values
are how many times a user has watched that particular movie. This matrix would
have tens of thousands of columns and millions of rows! However, since most users
do not watch most movies, the vast majority of elements would be zero.
Sparse matrices only store nonzero elements and assume all other values will be zero,
leading to significant computational savings. In our solution, we created a NumPy
array with two nonzero values, then converted it into a sparse matrix. If we view the
sparse matrix we can see that only the nonzero values are stored:
# View sparse matrix
print(matrix_sparse)
(1, 1)
(2, 0)
1
3
1.3 Creating a Sparse Matrix
|
3
There are a number of types of sparse matrices. However, in compressed sparse row
(CSR) matrices, (1, 1) and (2, 0) represent the (zero-indexed) indices of the nonzero values 1 and 3, respectively. For example, the element 1 is in the second row and
second column. We can see the advantage of sparse matrices if we create a much
larger matrix with many more zero elements and then compare this larger matrix
with our original sparse matrix:
# Create larger matrix
matrix_large = np.array([[0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
[0, 1, 0, 0, 0, 0, 0, 0, 0, 0],
[3, 0, 0, 0, 0, 0, 0, 0, 0, 0]])
# Create compressed sparse row (CSR) matrix
matrix_large_sparse = sparse.csr_matrix(matrix_large)
# View original sparse matrix
print(matrix_sparse)
(1, 1)
(2, 0)
1
3
# View larger sparse matrix
print(matrix_large_sparse)
(1, 1)
(2, 0)
1
3
As we can see, despite the fact that we added many more zero elements in the larger
matrix, its sparse representation is exactly the same as our original sparse matrix.
That is, the addition of zero elements did not change the size of the sparse matrix.
As mentioned, there are many different types of sparse matrices, such as compressed
sparse column, list of lists, and dictionary of keys. While an explanation of the differ‐
ent types and their implications is outside the scope of this book, it is worth noting
that while there is no “best” sparse matrix type, there are meaningful differences
between them and we should be conscious about why we are choosing one type over
another.
See Also
• Sparse matrices, SciPy documentation
• 101 Ways to Store a Sparse Matrix
1.4 Selecting Elements
Problem
You need to select one or more elements in a vector or matrix.
4
|
Chapter 1: Vectors, Matrices, and Arrays
Solution
NumPy’s arrays make that easy:
# Load library
import numpy as np
# Create row vector
vector = np.array([1, 2, 3, 4, 5, 6])
# Create matrix
matrix = np.array([[1, 2, 3],
[4, 5, 6],
[7, 8, 9]])
# Select third element of vector
vector[2]
3
# Select second row, second column
matrix[1,1]
5
Discussion
Like most things in Python, NumPy arrays are zero-indexed, meaning that the index
of the first element is 0, not 1. With that caveat, NumPy offers a wide variety of meth‐
ods for selecting (i.e., indexing and slicing) elements or groups of elements in arrays:
# Select all elements of a vector
vector[:]
array([1, 2, 3, 4, 5, 6])
# Select everything up to and including the third element
vector[:3]
array([1, 2, 3])
# Select everything after the third element
vector[3:]
array([4, 5, 6])
# Select the last element
vector[-1]
6
# Select the first two rows and all columns of a matrix
matrix[:2,:]
array([[1, 2, 3],
[4, 5, 6]])
1.4 Selecting Elements
|
5
# Select all rows and the second column
matrix[:,1:2]
array([[2],
[5],
[8]])
1.5 Describing a Matrix
Problem
You want to describe the shape, size, and dimensions of the matrix.
Solution
Use shape, size, and ndim:
# Load library
import numpy as np
# Create matrix
matrix = np.array([[1, 2, 3, 4],
[5, 6, 7, 8],
[9, 10, 11, 12]])
# View number of rows and columns
matrix.shape
(3, 4)
# View number of elements (rows * columns)
matrix.size
12
# View number of dimensions
matrix.ndim
2
Discussion
This might seem basic (and it is); however, time and again it will be valuable to check
the shape and size of an array both for further calculations and simply as a gut check
after some operation.
1.6 Applying Operations to Elements
Problem
You want to apply some function to multiple elements in an array.
6
|
Chapter 1: Vectors, Matrices, and Arrays
Solution
Use NumPy’s vectorize:
# Load library
import numpy as np
# Create matrix
matrix = np.array([[1, 2, 3],
[4, 5, 6],
[7, 8, 9]])
# Create function that adds 100 to something
add_100 = lambda i: i + 100
# Create vectorized function
vectorized_add_100 = np.vectorize(add_100)
# Apply function to all elements in matrix
vectorized_add_100(matrix)
array([[101, 102, 103],
[104, 105, 106],
[107, 108, 109]])
Discussion
NumPy’s vectorize class converts a function into a function that can apply to all ele‐
ments in an array or slice of an array. It’s worth noting that vectorize is essentially a
for loop over the elements and does not increase performance. Furthermore, NumPy
arrays allow us to perform operations between arrays even if their dimensions are not
the same (a process called broadcasting). For example, we can create a much simpler
version of our solution using broadcasting:
# Add 100 to all elements
matrix + 100
array([[101, 102, 103],
[104, 105, 106],
[107, 108, 109]])
1.7 Finding the Maximum and Minimum Values
Problem
You need to find the maximum or minimum value in an array.
Solution
Use NumPy’s max and min:
1.7 Finding the Maximum and Minimum Values
|
7
# Load library
import numpy as np
# Create matrix
matrix = np.array([[1, 2, 3],
[4, 5, 6],
[7, 8, 9]])
# Return maximum element
np.max(matrix)
9
# Return minimum element
np.min(matrix)
1
Discussion
Often we want to know the maximum and minimum value in an array or subset of an
array. This can be accomplished with the max and min methods. Using the axis
parameter we can also apply the operation along a certain axis:
# Find maximum element in each column
np.max(matrix, axis=0)
array([7, 8, 9])
# Find maximum element in each row
np.max(matrix, axis=1)
array([3, 6, 9])
1.8 Calculating the Average, Variance, and Standard
Deviation
Problem
You want to calculate some descriptive statistics about an array.
Solution
Use NumPy’s mean, var, and std:
# Load library
import numpy as np
# Create matrix
matrix = np.array([[1, 2, 3],
[4, 5, 6],
[7, 8, 9]])
8
|
Chapter 1: Vectors, Matrices, and Arrays
# Return mean
np.mean(matrix)
5.0
# Return variance
np.var(matrix)
6.666666666666667
# Return standard deviation
np.std(matrix)
2.5819888974716112
Discussion
Just like with max and min, we can easily get descriptive statistics about the whole
matrix or do calculations along a single axis:
# Find the mean value in each column
np.mean(matrix, axis=0)
array([ 4.,
5.,
6.])
1.9 Reshaping Arrays
Problem
You want to change the shape (number of rows and columns) of an array without
changing the element values.
Solution
Use NumPy’s reshape:
# Load library
import numpy as np
# Create 4x3 matrix
matrix = np.array([[1, 2, 3],
[4, 5, 6],
[7, 8, 9],
[10, 11, 12]])
# Reshape matrix into 2x6 matrix
matrix.reshape(2, 6)
array([[ 1,
[ 7,
2,
8,
3, 4, 5, 6],
9, 10, 11, 12]])
1.9 Reshaping Arrays
|
9
Discussion
reshape allows us to restructure an array so that we maintain the same data but it is
organized as a different number of rows and columns. The only requirement is that
the shape of the original and new matrix contain the same number of elements (i.e.,
the same size). We can see the size of a matrix using size:
matrix.size
12
One useful argument in reshape is -1, which effectively means “as many as needed,”
so reshape(-1, 1) means one row and as many columns as needed:
matrix.reshape(1, -1)
array([[ 1,
2,
3,
4,
5,
6,
7,
8,
9, 10, 11, 12]])
Finally, if we provide one integer, reshape will return a 1D array of that length:
matrix.reshape(12)
array([ 1,
2,
3,
4,
5,
6,
7,
8,
9, 10, 11, 12])
1.10 Transposing a Vector or Matrix
Problem
You need to transpose a vector or matrix.
Solution
Use the T method:
# Load library
import numpy as np
# Create matrix
matrix = np.array([[1, 2, 3],
[4, 5, 6],
[7, 8, 9]])
# Transpose matrix
matrix.T
array([[1, 4, 7],
[2, 5, 8],
[3, 6, 9]])
10
|
Chapter 1: Vectors, Matrices, and Arrays
Discussion
Transposing is a common operation in linear algebra where the column and row
indices of each element are swapped. One nuanced point that is typically overlooked
outside of a linear algebra class is that, technically, a vector cannot be transposed
because it is just a collection of values:
# Transpose vector
np.array([1, 2, 3, 4, 5, 6]).T
array([1, 2, 3, 4, 5, 6])
However, it is common to refer to transposing a vector as converting a row vector to
a column vector (notice the second pair of brackets) or vice versa:
# Tranpose row vector
np.array([[1, 2, 3, 4, 5, 6]]).T
array([[1],
[2],
[3],
[4],
[5],
[6]])
1.11 Flattening a Matrix
Problem
You need to transform a matrix into a one-dimensional array.
Solution
Use flatten:
# Load library
import numpy as np
# Create matrix
matrix = np.array([[1, 2, 3],
[4, 5, 6],
[7, 8, 9]])
# Flatten matrix
matrix.flatten()
array([1, 2, 3, 4, 5, 6, 7, 8, 9])
1.11 Flattening a Matrix
|
11
Discussion
flatten is a simple method to transform a matrix into a one-dimensional array.
Alternatively, we can use reshape to create a row vector:
matrix.reshape(1, -1)
array([[1, 2, 3, 4, 5, 6, 7, 8, 9]])
1.12 Finding the Rank of a Matrix
Problem
You need to know the rank of a matrix.
Solution
Use NumPy’s linear algebra method matrix_rank:
# Load library
import numpy as np
# Create matrix
matrix = np.array([[1, 1, 1],
[1, 1, 10],
[1, 1, 15]])
# Return matrix rank
np.linalg.matrix_rank(matrix)
2
Discussion
The rank of a matrix is the dimensions of the vector space spanned by its columns or
rows. Finding the rank of a matrix is easy in NumPy thanks to matrix_rank.
See Also
• The Rank of a Matrix, CliffsNotes
1.13 Calculating the Determinant
Problem
You need to know the determinant of a matrix.
12
|
Chapter 1: Vectors, Matrices, and Arrays
Solution
Use NumPy’s linear algebra method det:
# Load library
import numpy as np
# Create matrix
matrix = np.array([[1, 2, 3],
[2, 4, 6],
[3, 8, 9]])
# Return determinant of matrix
np.linalg.det(matrix)
0.0
Discussion
It can sometimes be useful to calculate the determinant of a matrix. NumPy makes
this easy with det.
See Also
• The determinant | Essence of linear algebra, chapter 5, 3Blue1Brown
• Determinant, Wolfram MathWorld
1.14 Getting the Diagonal of a Matrix
Problem
You need to get the diagonal elements of a matrix.
Solution
Use diagonal:
# Load library
import numpy as np
# Create matrix
matrix = np.array([[1, 2, 3],
[2, 4, 6],
[3, 8, 9]])
# Return diagonal elements
matrix.diagonal()
1.14 Getting the Diagonal of a Matrix
|
13
array([1, 4, 9])
Discussion
NumPy makes getting the diagonal elements of a matrix easy with diagonal. It is also
possible to get a diagonal off from the main diagonal by using the offset parameter:
# Return diagonal one above the main diagonal
matrix.diagonal(offset=1)
array([2, 6])
# Return diagonal one below the main diagonal
matrix.diagonal(offset=-1)
array([2, 8])
1.15 Calculating the Trace of a Matrix
Problem
You need to calculate the trace of a matrix.
Solution
Use trace:
# Load library
import numpy as np
# Create matrix
matrix = np.array([[1, 2, 3],
[2, 4, 6],
[3, 8, 9]])
# Return trace
matrix.trace()
14
Discussion
The trace of a matrix is the sum of the diagonal elements and is often used under the
hood in machine learning methods. Given a NumPy multidimensional array, we can
calculate the trace using trace. We can also return the diagonal of a matrix and calcu‐
late its sum:
# Return diagonal and sum elements
sum(matrix.diagonal())
14
14
|
Chapter 1: Vectors, Matrices, and Arrays
See Also
• The Trace of a Square Matrix
1.16 Finding Eigenvalues and Eigenvectors
Problem
You need to find the eigenvalues and eigenvectors of a square matrix.
Solution
Use NumPy’s linalg.eig:
# Load library
import numpy as np
# Create matrix
matrix = np.array([[1, -1, 3],
[1, 1, 6],
[3, 8, 9]])
# Calculate eigenvalues and eigenvectors
eigenvalues, eigenvectors = np.linalg.eig(matrix)
# View eigenvalues
eigenvalues
array([ 13.55075847,
0.74003145,
-3.29078992])
# View eigenvectors
eigenvectors
array([[-0.17622017, -0.96677403, -0.53373322],
[-0.435951 , 0.2053623 , -0.64324848],
[-0.88254925, 0.15223105, 0.54896288]])
Discussion
Eigenvectors are widely used in machine learning libraries. Intuitively, given a linear
transformation represented by a matrix, A, eigenvectors are vectors that, when that
transformation is applied, change only in scale (not direction). More formally:
Av = λv
where A is a square matrix, λ contains the eigenvalues and v contains the eigenvec‐
tors. In NumPy’s linear algebra toolset, eig lets us calculate the eigenvalues, and
eigenvectors of any square matrix.
1.16 Finding Eigenvalues and Eigenvectors
|
15
See Also
• Eigenvectors and Eigenvalues Explained Visually, Setosa.io
• Eigenvectors and eigenvalues | Essence of linear algebra, Chapter 10,
3Blue1Brown
1.17 Calculating Dot Products
Problem
You need to calculate the dot product of two vectors.
Solution
Use NumPy’s dot:
# Load library
import numpy as np
# Create two vectors
vector_a = np.array([1,2,3])
vector_b = np.array([4,5,6])
# Calculate dot product
np.dot(vector_a, vector_b)
32
Discussion
The dot product of two vectors, a and b, is defined as:
n
∑ aibi
i=1
where ai is the ith element of vector a. We can use NumPy’s dot class to calculate the
dot product. Alternatively, in Python 3.5+ we can use the new @ operator:
# Calculate dot product
vector_a @ vector_b
32
See Also
• Vector dot product and vector length, Khan Academy
16
|
Chapter 1: Vectors, Matrices, and Arrays
• Dot Product, Paul’s Online Math Notes
1.18 Adding and Subtracting Matrices
Problem
You want to add or subtract two matrices.
Solution
Use NumPy’s add and subtract:
# Load library
import numpy as np
# Create matrix
matrix_a = np.array([[1, 1, 1],
[1, 1, 1],
[1, 1, 2]])
# Create matrix
matrix_b = np.array([[1, 3, 1],
[1, 3, 1],
[1, 3, 8]])
# Add two matrices
np.add(matrix_a, matrix_b)
array([[ 2,
[ 2,
[ 2,
4, 2],
4, 2],
4, 10]])
# Subtract two matrices
np.subtract(matrix_a, matrix_b)
array([[ 0, -2, 0],
[ 0, -2, 0],
[ 0, -2, -6]])
Discussion
Alternatively, we can simply use the + and - operators:
# Add two matrices
matrix_a + matrix_b
array([[ 2,
[ 2,
[ 2,
4, 2],
4, 2],
4, 10]])
1.18 Adding and Subtracting Matrices
|
17
1.19 Multiplying Matrices
Problem
You want to multiply two matrices.
Solution
Use NumPy’s dot:
# Load library
import numpy as np
# Create matrix
matrix_a = np.array([[1, 1],
[1, 2]])
# Create matrix
matrix_b = np.array([[1, 3],
[1, 2]])
# Multiply two matrices
np.dot(matrix_a, matrix_b)
array([[2, 5],
[3, 7]])
Discussion
Alternatively, in Python 3.5+ we can use the @ operator:
# Multiply two matrices
matrix_a @ matrix_b
array([[2, 5],
[3, 7]])
If we want to do element-wise multiplication, we can use the * operator:
# Multiply two matrices element-wise
matrix_a * matrix_b
array([[1, 3],
[1, 4]])
See Also
• Array vs. Matrix Operations, MathWorks
18
|
Chapter 1: Vectors, Matrices, and Arrays
1.20 Inverting a Matrix
Problem
You want to calculate the inverse of a square matrix.
Solution
Use NumPy’s linear algebra inv method:
# Load library
import numpy as np
# Create matrix
matrix = np.array([[1, 4],
[2, 5]])
# Calculate inverse of matrix
np.linalg.inv(matrix)
array([[-1.66666667, 1.33333333],
[ 0.66666667, -0.33333333]])
Discussion
The inverse of a square matrix, A, is a second matrix A–1, such that:
AA−1 = I
where I is the identity matrix. In NumPy we can use linalg.inv to calculate A–1 if it
exists. To see this in action, we can multiply a matrix by its inverse and the result is
the identity matrix:
# Multiply matrix and its inverse
matrix @ np.linalg.inv(matrix)
array([[ 1.,
[ 0.,
0.],
1.]])
See Also
• Inverse of a Matrix
1.20 Inverting a Matrix
|
19
1.21 Generating Random Values
Problem
You want to generate pseudorandom values.
Solution
Use NumPy’s random:
# Load library
import numpy as np
# Set seed
np.random.seed(0)
# Generate three random floats between 0.0 and 1.0
np.random.random(3)
array([ 0.5488135 ,
0.71518937,
0.60276338])
Discussion
NumPy offers a wide variety of means to generate random numbers, many more than
can be covered here. In our solution we generated floats; however, it is also common
to generate integers:
# Generate three random integers between 1 and 10
np.random.randint(0, 11, 3)
array([3, 7, 9])
Alternatively, we can generate numbers by drawing them from a distribution:
# Draw three numbers from a normal distribution with mean 0.0
# and standard deviation of 1.0
np.random.normal(0.0, 1.0, 3)
array([-1.42232584,
1.52006949, -0.29139398])
# Draw three numbers from a logistic distribution with mean 0.0 and scale of 1.0
np.random.logistic(0.0, 1.0, 3)
array([-0.98118713, -0.08939902,
1.46416405])
# Draw three numbers greater than or equal to 1.0 and less than 2.0
np.random.uniform(1.0, 2.0, 3)
array([ 1.47997717,
1.3927848 ,
1.83607876])
Finally, it can sometimes be useful to return the same random numbers multiple
times to get predictable, repeatable results. We can do this by setting the “seed” (an
integer) of the pseudorandom generator. Random processes with the same seed will
20
|
Chapter 1: Vectors, Matrices, and Arrays
always produce the same output. We will use seeds throughout this book so that the
code you see in the book and the code you run on your computer produces the same
results.
1.21 Generating Random Values
|
21
CHAPTER 2
Loading Data
2.0 Introduction
The first step in any machine learning endeavor is to get the raw data into our system.
The raw data might be a logfile, dataset file, or database. Furthermore, often we will
want to retrieve data from multiple sources. The recipes in this chapter look at meth‐
ods of loading data from a variety of sources, including CSV files and SQL databases.
We also cover methods of generating simulated data with desirable properties for
experimentation. Finally, while there are many ways to load data in the Python eco‐
system, we will focus on using the pandas library’s extensive set of methods for load‐
ing external data, and using scikit-learn—an open source machine learning library in
Python—for generating simulated data.
2.1 Loading a Sample Dataset
Problem
You want to load a preexisting sample dataset.
Solution
scikit-learn comes with a number of popular datasets for you to use:
# Load scikit-learn's datasets
from sklearn import datasets
# Load digits dataset
digits = datasets.load_digits()
# Create features matrix
features = digits.data
23
# Create target vector
target = digits.target
# View first observation
features[0]
array([
0.,
15.,
8.,
5.,
1.,
0.,
0.,
10.,
0.,
8.,
12.,
0.,
5.,
15.,
0.,
0.,
7.,
0.,
13.,
5.,
4.,
0.,
0.,
6.,
9.,
0.,
12.,
9.,
0.,
13.,
1.,
0.,
0.,
8.,
2.,
10.,
0.,
3.,
0.,
0.,
14.,
0.,
0.,
15.,
8.,
0.,
5.,
0.,
0.,
0.,
2.,
0.,
8.,
0.,
4., 11.,
10., 12.,
0.])
13.,
11.,
0.,
0.,
0.,
Discussion
Often we do not want to go through the work of loading, transforming, and cleaning
a real-world dataset before we can explore some machine learning algorithm or
method. Luckily, scikit-learn comes with some common datasets we can quickly load.
These datasets are often called “toy” datasets because they are far smaller and cleaner
than a dataset we would see in the real world. Some popular sample datasets in scikitlearn are:
load_boston
Contains 503 observations on Boston housing prices. It is a good dataset for
exploring regression algorithms.
load_iris
Contains 150 observations on the measurements of Iris flowers. It is a good data‐
set for exploring classification algorithms.
load_digits
Contains 1,797 observations from images of handwritten digits. It is a good data‐
set for teaching image classification.
See Also
• scikit-learn toy datasets
• The Digit Dataset
2.2 Creating a Simulated Dataset
Problem
You need to generate a dataset of simulated data.
24
|
Chapter 2: Loading Data
Solution
scikit-learn offers many methods for creating simulated data. Of those, three methods
are particularly useful.
When we want a dataset designed to be used with linear regression, make_regression
is a good choice:
# Load library
from sklearn.datasets import make_regression
# Generate features matrix, target vector, and the true coefficients
features, target, coefficients = make_regression(n_samples = 100,
n_features = 3,
n_informative = 3,
n_targets = 1,
noise = 0.0,
coef = True,
random_state = 1)
# View feature matrix and target vector
print('Feature Matrix\n', features[:3])
print('Target Vector\n', target[:3])
Feature Matrix
[[ 1.29322588 -0.61736206 -0.11044703]
[-2.793085
0.36633201 1.93752881]
[ 0.80186103 -0.18656977 0.0465673 ]]
Target Vector
[-10.37865986 25.5124503
19.67705609]
If we are interested in creating a simulated dataset for classification, we can use
make_classification:
# Load library
from sklearn.datasets import make_classification
# Generate features matrix and target vector
features, target = make_classification(n_samples = 100,
n_features = 3,
n_informative = 3,
n_redundant = 0,
n_classes = 2,
weights = [.25, .75],
random_state = 1)
# View feature matrix and target vector
print('Feature Matrix\n', features[:3])
print('Target Vector\n', target[:3])
Feature Matrix
[[ 1.06354768 -1.42632219 1.02163151]
[ 0.23156977 1.49535261 0.33251578]
2.2 Creating a Simulated Dataset
|
25
[ 0.15972951
Target Vector
[1 0 0]
0.83533515 -0.40869554]]
Finally, if we want a dataset designed to work well with clustering techniques, scikitlearn offers make_blobs:
# Load library
from sklearn.datasets import make_blobs
# Generate feature matrix and target vector
features, target = make_blobs(n_samples = 100,
n_features = 2,
centers = 3,
cluster_std = 0.5,
shuffle = True,
random_state = 1)
# View feature matrix and target vector
print('Feature Matrix\n', features[:3])
print('Target Vector\n', target[:3])
Feature Matrix
[[ -1.22685609
3.25572052]
[ -9.57463218 -4.38310652]
[-10.71976941 -4.20558148]]
Target Vector
[0 1 1]
Discussion
As might be apparent from the solutions, make_regression returns a feature matrix
of float values and a target vector of float values, while make_classification and
make_blobs return a feature matrix of float values and a target vector of integers rep‐
resenting membership in a class.
scikit-learn’s simulated datasets offer extensive options to control the type of data
generated. scikit-learn’s documentation contains a full description of all the parame‐
ters, but a few are worth noting.
In make_regression and make_classification, n_informative determines the
number of features that are used to generate the target vector. If n_informative is less
than the total number of features (n_features), the resulting dataset will have redun‐
dant features that can be identified through feature selection techniques.
In addition, make_classification contains a weights parameter that allows us to
simulate datasets with imbalanced classes. For example, weights = [.25, .75]
would return a dataset with 25% of observations belonging to one class and 75% of
observations belonging to a second class.
26
|
Chapter 2: Loading Data
For make_blobs, the centers parameter determines the number of clusters generated.
Using the matplotlib visualization library, we can visualize the clusters generated by
make_blobs:
# Load library
import matplotlib.pyplot as plt
# View scatterplot
plt.scatter(features[:,0], features[:,1], c=target)
plt.show()
See Also
• make_regression documentation
• make_classification documentation
• make_blobs documentation
2.3 Loading a CSV File
Problem
You need to import a comma-separated values (CSV) file.
2.3 Loading a CSV File
|
27
Solution
Use the pandas library’s read_csv to load a local or hosted CSV file:
# Load library
import pandas as pd
# Create URL
url = 'https://tinyurl.com/simulated_data'
# Load dataset
dataframe = pd.read_csv(url)
# View first two rows
dataframe.head(2)
integer datetime
category
0 5
2015-01-01 00:00:00 0
1 5
2015-01-01 00:00:01 0
Discussion
There are two things to note about loading CSV files. First, it is often useful to take a
quick look at the contents of the file before loading. It can be very helpful to see how a
dataset is structured beforehand and what parameters we need to set to load in the
file. Second, read_csv has over 30 parameters and therefore the documentation can
be daunting. Fortunately, those parameters are mostly there to allow it to handle a
wide variety of CSV formats. For example, CSV files get their names from the fact
that the values are literally separated by commas (e.g., one row might be
2,"2015-01-01 00:00:00",0); however, it is common for “CSV” files to use other
characters as separators, like tabs. pandas’ sep parameter allows us to define the
delimiter used in the file. Although it is not always the case, a common formatting
issue with CSV files is that the first line of the file is used to define column headers
(e.g., integer, datetime, category in our solution). The header parameter allows
us to specify if or where a header row exists. If a header row does not exist, we set
header=None.
2.4 Loading an Excel File
Problem
You need to import an Excel spreadsheet.
Solution
Use the pandas library’s read_excel to load an Excel spreadsheet:
28
|
Chapter 2: Loading Data
# Load library
import pandas as pd
# Create URL
url = 'https://tinyurl.com/simulated_excel'
# Load data
dataframe = pd.read_excel(url, sheetname=0, header=1)
# View the first two rows
dataframe.head(2)
5 2015-01-01 00:00:00 0
0 5 2015-01-01 00:00:01 0
1 9 2015-01-01 00:00:02 0
Discussion
This solution is similar to our solution for reading CSV files. The main difference is
the additional parameter, sheetname, that specifies which sheet in the Excel file we
wish to load. sheetname can accept both strings containing the name of the sheet and
integers pointing to sheet positions (zero-indexed). If we need to load multiple sheets,
include them as a list. For example, sheetname=[0,1,2, "Monthly Sales"] will
return a dictionary of pandas DataFrames containing the first, second, and third
sheets and the sheet named Monthly Sales.
2.5 Loading a JSON File
Problem
You need to load a JSON file for data preprocessing.
Solution
The pandas library provides read_json to convert a JSON file into a pandas object:
# Load library
import pandas as pd
# Create URL
url = 'https://tinyurl.com/simulated_json'
# Load data
dataframe = pd.read_json(url, orient='columns')
# View the first two rows
dataframe.head(2)
2.5 Loading a JSON File
|
29
category
datetime
integer
0
0
2015-01-01 00:00:00
5
1
0
2015-01-01 00:00:01
5
Discussion
Importing JSON files into pandas is similar to the last few recipes we have seen. The
key difference is the orient parameter, which indicates to pandas how the JSON file
is structured. However, it might take some experimenting to figure out which argu‐
ment ( split, records, index, columns, and values) is the right one. Another helpful
tool pandas offers is json_normalize, which can help convert semistructured JSON
data into a pandas DataFrame.
See Also
• json_normalize documentation
2.6 Querying a SQL Database
Problem
You need to load data from a database using the structured query language (SQL).
Solution
pandas’ read_sql_query allows us to make a SQL query to a database and load it:
# Load libraries
import pandas as pd
from sqlalchemy import create_engine
# Create a connection to the database
database_connection = create_engine('sqlite:///sample.db')
# Load data
dataframe = pd.read_sql_query('SELECT * FROM data', database_connection)
# View first two rows
dataframe.head(2)
30
|
Chapter 2: Loading Data
first_name last_name age preTestScore postTestScore
0 Jason
Miller
42 4
25
1 Molly
Jacobson
52 24
94
Discussion
Out of all of the recipes presented in this chapter, this recipe is probably the one we
will use most in the real world. SQL is the lingua franca for pulling data from data‐
bases. In this recipe we first use create_engine to define a connection to a SQL data‐
base engine called SQLite. Next we use pandas’ read_sql_query to query that data‐
base using SQL and put the results in a DataFrame.
SQL is a language in its own right and, while beyond the scope of this book, it is cer‐
tainly worth knowing for anyone wanting to learn machine learning. Our SQL query,
SELECT * FROM data, asks the database to give us all columns (*) from the table
called data.
See Also
• SQLite
• W3Schools SQL Tutorial
2.6 Querying a SQL Database
|
31
CHAPTER 3
Data Wrangling
3.0 Introduction
Data wrangling is a broad term used, often informally, to describe the process of
transforming raw data to a clean and organized format ready for use. For us, data
wrangling is only one step in preprocessing our data, but it is an important step.
The most common data structure used to “wrangle” data is the data frame, which can
be both intuitive and incredibly versatile. Data frames are tabular, meaning that they
are based on rows and columns like you would see in a spreadsheet. Here is a data
frame created from data about passengers on the Titanic:
# Load library
import pandas as pd
# Create URL
url = 'https://tinyurl.com/titanic-csv'
# Load data as a dataframe
dataframe = pd.read_csv(url)
# Show first 5 rows
dataframe.head(5)
0
1
2
3
4
Name
Allen, Miss Elisabeth Walton
Allison, Miss Helen Loraine
Allison, Mr Hudson Joshua Creighton
Allison, Mrs Hudson JC (Bessie Waldo Daniels)
Allison, Master Hudson Trevor
PClass
1st
1st
1st
1st
1st
Age
29.00
2.00
30.00
25.00
0.92
Sex
female
female
male
female
male
Survived
1
0
0
0
1
SexCode
1
1
0
1
0
33
There are three important things to notice in this data frame.
First, in a data frame each row corresponds to one observation (e.g., a passenger) and
each column corresponds to one feature (gender, age, etc.). For example, by looking
at the first observation we can see that Miss Elisabeth Walton Allen stayed in first
class, was 29 years old, was female, and survived the disaster.
Second, each column contains a name (e.g., Name, PClass, Age) and each row contains
an index number (e.g., 0 for the lucky Miss Elisabeth Walton Allen). We will use these
to select and manipulate observations and features.
Third, two columns, Sex and SexCode, contain the same information in different for‐
mats. In Sex, a woman is indicated by the string female, while in SexCode, a woman
is indicated by using the integer 1. We will want all our features to be unique, and
therefore we will need to remove one of these columns.
In this chapter, we will cover a wide variety of techniques to manipulate data frames
using the pandas library with the goal of creating a clean, well-structured set of obser‐
vations for further preprocessing.
3.1 Creating a Data Frame
Problem
You want to create a new data frame.
Solution
pandas has many methods of creating a new DataFrame object. One easy method is
to create an empty data frame using DataFrame and then define each column sepa‐
rately:
# Load library
import pandas as pd
# Create DataFrame
dataframe = pd.DataFrame()
# Add columns
dataframe['Name'] = ['Jacky Jackson', 'Steven Stevenson']
dataframe['Age'] = [38, 25]
dataframe['Driver'] = [True, False]
# Show DataFrame
dataframe
34
|
Chapter 3: Data Wrangling
Name
Age Driver
0 Jacky Jackson
38 True
1 Steven Stevenson 25 False
Alternatively, once we have created a DataFrame object, we can append new rows to
the bottom:
# Create row
new_person = pd.Series(['Molly Mooney', 40, True], index=['Name','Age','Driver'])
# Append row
dataframe.append(new_person, ignore_index=True)
Name
Age
0 Jacky Jackson
38
1 Steven Stevenson 25
2 Molly Mooney
40
Driver
True
False
True
Discussion
pandas offers what can feel like an infinite number of ways to create a DataFrame. In
the real world, creating an empty DataFrame and then populating it will almost never
happen. Instead, our DataFrames will be created from real data we have loading from
other sources (e.g., a CSV file or database).
3.2 Describing the Data
Problem
You want to view some characteristics of a DataFrame.
Solution
One of the easiest things we can do after loading the data is view the first few rows
using head:
# Load library
import pandas as pd
# Create URL
url = 'https://tinyurl.com/titanic-csv'
# Load data
dataframe = pd.read_csv(url)
# Show two rows
dataframe.head(2)
3.2 Describing the Data
|
35
Name
PClass Age Sex
Survived SexCode
0 Allen, Miss Elisabeth Walton 1st
29.0 female 1
1
1 Allison, Miss Helen Loraine 1st
2.0 female 0
1
We can also take a look at the number of rows and columns:
# Show dimensions
dataframe.shape
(1313, 6)
Additionally, we can get descriptive statistics for any numeric columns using
describe:
# Show statistics
dataframe.describe()
Age
count 756.000000
mean 30.397989
std
14.259049
min 0.170000
25% 21.000000
50% 28.000000
75% 39.000000
max 71.000000
Survived
1313.000000
0.342727
0.474802
0.000000
0.000000
0.000000
1.000000
1.000000
SexCode
1313.000000
0.351866
0.477734
0.000000
0.000000
0.000000
1.000000
1.000000
Discussion
After we load some data, it is a good idea to understand how it is structured and what
kind of information it contains. Ideally, we would view the full data directly. But with
most real-world cases, the data could have thousands to hundreds of thousands to
millions of rows and columns. Instead, we have to rely on pulling samples to view
small slices and calculating summary statistics of the data.
In our solution, we are using a toy dataset of the passengers of the Titanic on her last
voyage. Using head we can take a look at the first few rows (five by default) of the
data. Alternatively, we can use tail to view the last few rows. With shape we can see
how many rows and columns our DataFrame contains. And finally, with describe we
can see some basic descriptive statistics for any numerical column.
It is worth noting that summary statistics do not always tell the full story. For exam‐
ple, pandas treats the columns Survived and SexCode as numeric columns because
they contain 1s and 0s. However, in this case the numerical values represent cate‐
gories. For example, if Survived equals 1, it indicates that the passenger survived the
disaster. For this reason, some of the summary statistics provided don’t make sense,
36
|
Chapter 3: Data Wrangling
such as the standard deviation of the SexCode column (an indicator of the passenger’s
gender).
3.3 Navigating DataFrames
Problem
You need to select individual data or slices of a DataFrame.
Solution
Use loc or iloc to select one or more rows or values:
# Load library
import pandas as pd
# Create URL
url = 'https://tinyurl.com//titanic-csv'
# Load data
dataframe = pd.read_csv(url)
# Select first row
dataframe.iloc[0]
Name
Allen, Miss Elisabeth Walton
PClass
1st
Age
29
Sex
female
Survived
1
SexCode
1
Name: 0, dtype: object
We can use : to define a slice of rows we want, such as selecting the second, third,
and fourth rows:
# Select three rows
dataframe.iloc[1:4]
Name
1 Allison, Miss Helen Loraine
2 Allison, Mr Hudson Joshua Creighton
3 Allison, Mrs Hudson JC (Bessie Waldo Daniels)
PClass
1st
1st
1st
Age
2.0
30.0
25.0
Sex
female
male
female
Survived
0
0
0
SexCode
1
0
1
We can even use it to get all rows up to a point, such as all rows up to and including
the fourth row:
# Select three rows
dataframe.iloc[:4]
3.3 Navigating DataFrames
|
37
0
1
2
3
Name
Allen, Miss Elisabeth Walton
Allison, Miss Helen Loraine
Allison, Mr Hudson Joshua Creighton
Allison, Mrs Hudson JC (Bessie Waldo Daniels)
PClass
1st
1st
1st
1st
Age
29.0
2.0
30.0
25.0
Sex
female
female
male
female
Survived
1
0
0
0
SexCode
1
1
0
1
DataFrames do not need to be numerically indexed. We can set the index of a Data‐
Frame to any value where the value is unique to each row. For example, we can set the
index to be passenger names and then select rows using a name:
# Set index
dataframe = dataframe.set_index(dataframe['Name'])
# Show row
dataframe.loc['Allen, Miss Elisabeth Walton']
Name
Allen, Miss Elisabeth Walton
PClass
1st
Age
29
Sex
female
Survived
1
SexCode
1
Name: Allen, Miss Elisabeth Walton, dtype: object
Discussion
All rows in a pandas DataFrame have a unique index value. By default, this index is
an integer indicating the row position in the DataFrame; however, it does not have to
be. DataFrame indexes can be set to be unique alphanumeric strings or customer
numbers. To select individual rows and slices of rows, pandas provides two methods:
• loc is useful when the index of the DataFrame is a label (e.g., a string).
• iloc works by looking for the position in the DataFrame. For example, iloc[0]
will return the first row regardless of whether the index is an integer or a label.
It is useful to be comfortable with both loc and iloc since they will come up a lot
during data cleaning.
3.4 Selecting Rows Based on Conditionals
Problem
You want to select DataFrame rows based on some condition.
38
|
Chapter 3: Data Wrangling
Solution
This can be easily done in pandas. For example, if we wanted to select all the women
on the Titanic:
# Load library
import pandas as pd
# Create URL
url = 'https://tinyurl.com/titanic-csv'
# Load data
dataframe = pd.read_csv(url)
# Show top two rows where column 'sex' is 'female'
dataframe[dataframe['Sex'] == 'female'].head(2)
Name
PClass Age Sex
Survived SexCode
0 Allen, Miss Elisabeth Walton 1st
29.0 female 1
1
1 Allison, Miss Helen Loraine 1st
2.0 female 0
1
Take a second and look at the format of this solution. dataframe['Sex'] ==
'female' is our conditional statement; by wrapping that in dataframe[] we are tell‐
ing pandas to “select all the rows in the DataFrame where the value of data
frame['Sex'] is 'female'.
Multiple conditions are easy as well. For example, here we select all the rows where
the passenger is a female 65 or older:
# Filter rows
dataframe[(dataframe['Sex'] == 'female') & (dataframe['Age'] >= 65)]
Name
PClass Age Sex
Survived SexCode
73 Crosby, Mrs Edward Gifford (Catherine Elizabet... 1st
69.0 female 1
1
Discussion
Conditionally selecting and filtering data is one of the most common tasks in data
wrangling. You rarely want all the raw data from the source; instead, you are interes‐
ted in only some subsection of it. For example, you might only be interested in stores
in certain states or the records of patients over a certain age.
3.5 Replacing Values
Problem
You need to replace values in a DataFrame.
3.5 Replacing Values
|
39
Solution
pandas’ replace is an easy way to find and replace values. For example, we can
replace any instance of "female" in the Sex column with "Woman":
# Load library
import pandas as pd
# Create URL
url = 'https://tinyurl.com/titanic-csv'
# Load data
dataframe = pd.read_csv(url)
# Replace values, show two rows
dataframe['Sex'].replace("female", "Woman").head(2)
0
Woman
1
Woman
Name: Sex, dtype: object
We can also replace multiple values at the same time:
# Replace "female" and "male with "Woman" and "Man"
dataframe['Sex'].replace(["female", "male"], ["Woman", "Man"]).head(5)
0
Woman
1
Woman
2
Man
3
Woman
4
Man
Name: Sex, dtype: object
We can also find and replace across the entire DataFrame object by specifying the
whole data frame instead of a single column:
# Replace values, show two rows
dataframe.replace(1, "One").head(2)
Name
PClass Age Sex
Survived SexCode
0 Allen, Miss Elisabeth Walton 1st
29 female One
One
1 Allison, Miss Helen Loraine 1st
2
female 0
One
replace also accepts regular expressions:
# Replace values, show two rows
dataframe.replace(r"1st", "First", regex=True).head(2)
Name
PClass Age Sex
Survived SexCode
0 Allen, Miss Elisabeth Walton First
29.0 female 1
1
1 Allison, Miss Helen Loraine First
2.0 female 0
1
40
|
Chapter 3: Data Wrangling
Discussion
replace is a tool we use to replace values that is simple and yet has the powerful abil‐
ity to accept regular expressions.
3.6 Renaming Columns
Problem
You want to rename a column in a pandas DataFrame.
Solution
Rename columns using the rename method:
# Load library
import pandas as pd
# Create URL
url = 'https://tinyurl.com/titanic-csv'
# Load data
dataframe = pd.read_csv(url)
# Rename column, show two rows
dataframe.rename(columns={'PClass': 'Passenger Class'}).head(2)
Name
Passenger Class Age Sex
Survived SexCode
0 Allen, Miss Elisabeth Walton 1st
29.0 female 1
1
1 Allison, Miss Helen Loraine 1st
2.0 female 0
1
Notice that the rename method can accept a dictionary as a parameter. We can use the
dictionary to change multiple column names at once:
# Rename columns, show two rows
dataframe.rename(columns={'PClass': 'Passenger Class', 'Sex': 'Gender'}).head(2)
Name
Passenger Class Age Gender Survived SexCode
0 Allen, Miss Elisabeth Walton 1st
29.0 female 1
1
1 Allison, Miss Helen Loraine 1st
2.0 female 0
1
Discussion
Using rename with a dictionary as an argument to the columns parameter is my pre‐
ferred way to rename columns because it works with any number of columns. If we
want to rename all columns at once, this helpful snippet of code creates a dictionary
with the old column names as keys and empty strings as values:
3.6 Renaming Columns
|
41
# Load library
import collections
# Create dictionary
column_names = collections.defaultdict(str)
# Create keys
for name in dataframe.columns:
column_names[name]
# Show dictionary
column_names
defaultdict(str,
{'Age': '',
'Name': '',
'PClass': '',
'Sex': '',
'SexCode': '',
'Survived': ''})
3.7 Finding the Minimum, Maximum, Sum, Average, and
Count
Problem
You want to find the min, max, sum, average, or count of a numeric column.
Solution
pandas comes with some built-in methods for commonly used descriptive statistics:
# Load library
import pandas as pd
# Create URL
url = 'https://tinyurl.com/titanic-csv'
# Load data
dataframe = pd.read_csv(url)
# Calculate statistics
print('Maximum:', dataframe['Age'].max())
print('Minimum:', dataframe['Age'].min())
print('Mean:', dataframe['Age'].mean())
print('Sum:', dataframe['Age'].sum())
print('Count:', dataframe['Age'].count())
Maximum: 71.0
Minimum: 0.17
Mean: 30.397989417989415
42
|
Chapter 3: Data Wrangling
Sum: 22980.879999999997
Count: 756
Discussion
In addition to the statistics used in the solution, pandas offers variance (var), stan‐
dard deviation (std), kurtosis (kurt), skewness (skew), standard error of the mean
(sem), mode (mode), median (median), and a number of others.
Furthermore, we can also apply these methods to the whole DataFrame:
# Show counts
dataframe.count()
Name
1313
PClass
1313
Age
756
Sex
1313
Survived
1313
SexCode
1313
dtype: int64
3.8 Finding Unique Values
Problem
You want to select all unique values in a column.
Solution
Use unique to view an array of all unique values in a column:
# Load library
import pandas as pd
# Create URL
url = 'https://tinyurl.com/titanic-csv'
# Load data
dataframe = pd.read_csv(url)
# Select unique values
dataframe['Sex'].unique()
array(['female', 'male'], dtype=object)
Alternatively, value_counts will display all unique values with the number of times
each value appears:
# Show counts
dataframe['Sex'].value_counts()
3.8 Finding Unique Values
|
43
male
851
female
462
Name: Sex, dtype: int64
Discussion
Both unique and value_counts are useful for manipulating and exploring categorical
columns. Very often in categorical columns there will be classes that need to be han‐
dled in the data wrangling phase. For example, in the Titanic dataset, PClass is a col‐
umn indicating the class of a passenger’s ticket. There were three classes on the
Titanic; however, if we use value_counts we can see a problem:
# Show counts
dataframe['PClass'].value_counts()
3rd
711
1st
322
2nd
279
*
1
Name: PClass, dtype: int64
While almost all passengers belong to one of three classes as expected, a single pas‐
senger has the class *. There are a number of strategies for handling this type of issue,
which we will address in Chapter 5, but for now just realize that “extra” classes are
common in categorical data and should not be ignored.
Finally, if we simply want to count the number of unique values, we can use nunique:
# Show number of unique values
dataframe['PClass'].nunique()
4
3.9 Handling Missing Values
Problem
You want to select missing values in a DataFrame.
Solution
isnull and notnull return booleans indicating whether a value is missing:
# Load library
import pandas as pd
# Create URL
url = 'https://tinyurl.com/titanic-csv'
# Load data
44
|
Chapter 3: Data Wrangling
dataframe = pd.read_csv(url)
## Select missing values, show two rows
dataframe[dataframe['Age'].isnull()].head(2)
Name
PClass Age Sex
Survived SexCode
12 Aubert, Mrs Leontine Pauline 1st
NaN female 1
1
13 Barkworth, Mr Algernon H
1st
NaN male 1
0
Discussion
Missing values are a ubiquitous problem in data wrangling, yet many underestimate
the difficulty of working with missing data. pandas uses NumPy’s NaN (“Not A Num‐
ber”) value to denote missing values, but it is important to note that NaN is not fully
implemented natively in pandas. For example, if we wanted to replace all strings con‐
taining male with missing values, we return an error:
# Attempt to replace values with NaN
dataframe['Sex'] = dataframe['Sex'].replace('male', NaN)
--------------------------------------------------------------------------NameError
Traceback (most recent call last)
<ipython-input-7-5682d714f87d> in <module>()
1 # Attempt to replace values with NaN
----> 2 dataframe['Sex'] = dataframe['Sex'].replace('male', NaN)
NameError: name 'NaN' is not defined
---------------------------------------------------------------------------
To have full functionality with NaN we need to import the NumPy library first:
# Load library
import numpy as np
# Replace values with NaN
dataframe['Sex'] = dataframe['Sex'].replace('male', np.nan)
Oftentimes a dataset uses a specific value to denote a missing observation, such as
NONE, -999, or .. pandas’ read_csv includes a parameter allowing us to specify the
values used to indicate missing values:
# Load data, set missing values
dataframe = pd.read_csv(url, na_values=[np.nan, 'NONE', -999])
3.9 Handling Missing Values
|
45
3.10 Deleting a Column
Problem
You want to delete a column from your DataFrame.
Solution
The best way to delete a column is to use drop with the parameter axis=1 (i.e., the
column axis):
# Load library
import pandas as pd
# Create URL
url = 'https://tinyurl.com/titanic-csv'
# Load data
dataframe = pd.read_csv(url)
# Delete column
dataframe.drop('Age', axis=1).head(2)
Name
PClass Sex
Survived SexCode
0 Allen, Miss Elisabeth Walton 1st
female 1
1
1 Allison, Miss Helen Loraine 1st
female 0
1
You can also use a list of column names as the main argument to drop multiple col‐
umns at once:
# Drop columns
dataframe.drop(['Age', 'Sex'], axis=1).head(2)
Name
PClass Survived SexCode
0 Allen, Miss Elisabeth Walton 1st
1
1
1 Allison, Miss Helen Loraine 1st
0
1
If a column does not have a name (which can sometimes happen), you can drop it by
its column index using dataframe.columns:
# Drop column
dataframe.drop(dataframe.columns[1], axis=1).head(2)
Name
Age Sex
Survived SexCode
0 Allen, Miss Elisabeth Walton 29.0 female 1
1
1 Allison, Miss Helen Loraine 2.0 female 0
1
46
|
Chapter 3: Data Wrangling
Discussion
drop is the idiomatic method of deleting a column. An alternative method is del
dataframe['Age'], which works most of the time but is not recommended because
of how it is called within pandas (the details of which are outside the scope of this
book).
One habit I recommend learning is to never use pandas’ inplace=True argument.
Many pandas methods include an inplace parameter, which when True edits the
DataFrame directly. This can lead to problems in more complex data processing pipe‐
lines because we are treating the DataFrames as mutable objects (which they techni‐
cally are). I recommend treating DataFrames as immutable objects. For example:
# Create a new DataFrame
dataframe_name_dropped = dataframe.drop(dataframe.columns[0], axis=1)
In this example, we are not mutating the DataFrame dataframe but instead are mak‐
ing a new DataFrame that is an altered version of dataframe called data
frame_name_dropped. If you treat your DataFrames as immutable objects, you will
save yourself a lot of headaches down the road.
3.11 Deleting a Row
Problem
You want to delete one or more rows from a DataFrame.
Solution
Use a boolean condition to create a new DataFrame excluding the rows you want to
delete:
# Load library
import pandas as pd
# Create URL
url = 'https://tinyurl.com/titanic-csv'
# Load data
dataframe = pd.read_csv(url)
# Delete rows, show first two rows of output
dataframe[dataframe['Sex'] != 'male'].head(2)
Name
PClass Age Sex
Survived SexCode
0 Allen, Miss Elisabeth Walton 1st
29.0 female 1
1
1 Allison, Miss Helen Loraine 1st
2.0 female 0
1
3.11 Deleting a Row
|
47
Discussion
While technically you can use the drop method (for example, df.drop([0, 1],
axis=0) to drop the first two rows), a more practical method is simply to wrap a
boolean condition inside df[]. The reason is because we can use the power of condi‐
tionals to delete either a single row or (far more likely) many rows at once.
We can use boolean conditions to easily delete single rows by matching a unique
value:
# Delete row, show first two rows of output
dataframe[dataframe['Name'] != 'Allison, Miss Helen Loraine'].head(2)
Name
PClass Age Sex
Survived SexCode
0 Allen, Miss Elisabeth Walton
1st
29.0 female 1
1
2 Allison, Mr Hudson Joshua Creighton 1st
30.0 male 0
0
And we can even use it to delete a single row by row index:
# Delete row, show first two rows of output
dataframe[dataframe.index != 0].head(2)
Name
PClass Age Sex
Survived SexCode
1 Allison, Miss Helen Loraine
1st
2.0 female 0
1
2 Allison, Mr Hudson Joshua Creighton 1st
30.0 male 0
0
3.12 Dropping Duplicate Rows
Problem
You want to drop duplicate rows from your DataFrame.
Solution
Use drop_duplicates, but be mindful of the parameters:
# Load library
import pandas as pd
# Create URL
url = 'https://tinyurl.com/titanic-csv'
# Load data
dataframe = pd.read_csv(url)
# Drop duplicates, show first two rows of output
dataframe.drop_duplicates().head(2)
48
|
Chapter 3: Data Wrangling
Name
PClass Age Sex
Survived SexCode
0 Allen, Miss Elisabeth Walton 1st
29.0 female 1
1
1 Allison, Miss Helen Loraine 1st
2.0 female 0
1
Discussion
A keen reader will notice that the solution didn’t actually drop any rows:
# Show number of rows
print("Number Of Rows In The Original DataFrame:", len(dataframe))
print("Number Of Rows After Deduping:", len(dataframe.drop_duplicates()))
Number Of Rows In The Original DataFrame: 1313
Number Of Rows After Deduping: 1313
The reason is because drop_duplicates defaults to only dropping rows that match
perfectly across all columns. Under this condition, every row in our DataFrame, data
frame, is actually unique. However, often we want to consider only a subset of col‐
umns to check for duplicate rows. We can accomplish this using the subset parame‐
ter:
# Drop duplicates
dataframe.drop_duplicates(subset=['Sex'])
Name
PClass Age Sex
Survived SexCode
0 Allen, Miss Elisabeth Walton
1st
29.0 female 1
1
2 Allison, Mr Hudson Joshua Creighton 1st
30.0 male 0
0
Take a close look at the preceding output: we told drop_duplicates to only consider
any two rows with the same value for Sex to be duplicates and to drop them. Now we
are left with a DataFrame of only two rows: one man and one woman. You might be
asking why drop_duplicates decided to keep these two rows instead of two different
rows. The answer is that drop_duplicates defaults to keeping the first occurrence of
a duplicated row and dropping the rest. We can control this behavior using the keep
parameter:
# Drop duplicates
dataframe.drop_duplicates(subset=['Sex'], keep='last')
Name
PClass Age Sex
Survived SexCode
1307 Zabour, Miss Tamini 3rd
NaN female 0
1
1312 Zimmerman, Leo
3rd
29.0 male 0
0
A related method is duplicated, which returns a boolean series denoting if a row is a
duplicate or not. This is a good option if you don’t want to simply drop duplicates.
3.12 Dropping Duplicate Rows
|
49
3.13 Grouping Rows by Values
Problem
You want to group individual rows according to some shared value.
Solution
groupby is one of the most powerful features in pandas:
# Load library
import pandas as pd
# Create URL
url = 'https://tinyurl.com/titanic-csv'
# Load data
dataframe = pd.read_csv(url)
# Group rows by the values of the column 'Sex', calculate mean
# of each group
dataframe.groupby('Sex').mean()
Age
Survived SexCode
Sex
female 29.396424 0.666667 1.0
male 31.014338 0.166863 0.0
Discussion
groupby is where data wrangling really starts to take shape. It is very common to have
a DataFrame where each row is a person or an event and we want to group them
according to some criterion and then calculate a statistic. For example, you can imag‐
ine a DataFrame where each row is an individual sale at a national restaurant chain
and we want the total sales per restaurant. We can accomplish this by grouping rows
by individual resturants and then calculating the sum of each group.
Users new to groupby often write a line like this and are confused by what is returned:
# Group rows
dataframe.groupby('Sex')
<pandas.core.groupby.DataFrameGroupBy object at 0x10efacf28>
Why didn’t it return something more useful? The reason is because groupby needs to
be paired with some operation we want to apply to each group, such as calculating an
aggregate statistic (e.g., mean, median, sum). When talking about grouping we often
use shorthand and say “group by gender,” but that is incomplete. For grouping to be
50
|
Chapter 3: Data Wrangling
useful, we need to group by something and then apply a function to each of those
groups:
# Group rows, count rows
dataframe.groupby('Survived')['Name'].count()
Survived
0
863
1
450
Name: Name, dtype: int64
Notice Name added after the groupby? That is because particular summary statistics
are only meaningful to certain types of data. For example, while calculating the aver‐
age age by gender makes sense, calculating the total age by gender does not. In this
case we group the data into survived or not, then count the number of names (i.e.,
passengers) in each group.
We can also group by a first column, then group that grouping by a second column:
# Group rows, calculate mean
dataframe.groupby(['Sex','Survived'])['Age'].mean()
Sex
female
Survived
0
24.901408
1
30.867143
male
0
32.320780
1
25.951875
Name: Age, dtype: float64
3.14 Grouping Rows by Time
Problem
You need to group individual rows by time periods.
Solution
Use resample to group rows by chunks of time:
# Load libraries
import pandas as pd
import numpy as np
# Create date range
time_index = pd.date_range('06/06/2017', periods=100000, freq='30S')
# Create DataFrame
dataframe = pd.DataFrame(index=time_index)
# Create column of random values
dataframe['Sale_Amount'] = np.random.randint(1, 10, 100000)
3.14 Grouping Rows by Time
|
51
# Group rows by week, calculate sum per week
dataframe.resample('W').sum()
2017-06-11
2017-06-18
2017-06-25
2017-07-02
2017-07-09
2017-07-16
Sale_Amount
86423
101045
100867
100894
100438
10297
Discussion
Our standard Titanic dataset does not contain a datetime column, so for this recipe
we have generated a simple DataFrame where each row is an individual sale. For each
sale we know its date and time and its dollar amount (this data isn’t realistic because
every sale takes place precisely 30 seconds apart and is an exact dollar amount, but
for the sake of simplicity let us pretend).
The raw data looks like this:
# Show three rows
dataframe.head(3)
Sale_Amount
2017-06-06 00:00:00 7
2017-06-06 00:00:30 2
2017-06-06 00:01:00 7
Notice that the date and time of each sale is the index of the DataFrame; this is
because resample requires the index to be datetime-like values.
Using resample we can group the rows by a wide array of time periods (offsets) and
then we can calculate some statistic on each time group:
# Group by two weeks, calculate mean
dataframe.resample('2W').mean()
2017-06-11
2017-06-25
2017-07-09
2017-07-23
52
|
Sale_Amount
5.001331
5.007738
4.993353
4.950481
Chapter 3: Data Wrangling
# Group by month, count rows
dataframe.resample('M').count()
Sale_Amount
2017-06-30 72000
2017-07-31 28000
You might notice that in the two outputs the datetime index is a date despite the fact
that we are grouping by weeks and months, respectively. The reason is because by
default resample returns the label of the right “edge” (the last label) of the time group.
We can control this behavior using the label parameter:
# Group by month, count rows
dataframe.resample('M', label='left').count()
Sale_Amount
2017-05-31 72000
2017-06-30 28000
See Also
• List of pandas time offset aliases
3.15 Looping Over a Column
Problem
You want to iterate over every element in a column and apply some action.
Solution
You can treat a pandas column like any other sequence in Python:
# Load library
import pandas as pd
# Create URL
url = 'https://tinyurl.com/titanic-csv'
# Load data
dataframe = pd.read_csv(url)
# Print first two names uppercased
for name in dataframe['Name'][0:2]:
print(name.upper())
3.15 Looping Over a Column
|
53
ALLEN, MISS ELISABETH WALTON
ALLISON, MISS HELEN LORAINE
Discussion
In addition to loops (often called for loops), we can also use list comprehensions:
# Show first two names uppercased
[name.upper() for name in dataframe['Name'][0:2]]
['ALLEN, MISS ELISABETH WALTON', 'ALLISON, MISS HELEN LORAINE']
Despite the temptation to fall back on for loops, a more Pythonic solution would use
pandas’ apply method, described in the next recipe.
3.16 Applying a Function Over All Elements in a Column
Problem
You want to apply some function over all elements in a column.
Solution
Use apply to apply a built-in or custom function on every element in a column:
# Load library
import pandas as pd
# Create URL
url = 'https://tinyurl.com/titanic-csv'
# Load data
dataframe = pd.read_csv(url)
# Create function
def uppercase(x):
return x.upper()
# Apply function, show two rows
dataframe['Name'].apply(uppercase)[0:2]
0
ALLEN, MISS ELISABETH WALTON
1
ALLISON, MISS HELEN LORAINE
Name: Name, dtype: object
Discussion
apply is a great way to do data cleaning and wrangling. It is common to write a func‐
tion to perform some useful operation (separate first and last names, convert strings
to floats, etc.) and then map that function to every element in a column.
54
|
Chapter 3: Data Wrangling
3.17 Applying a Function to Groups
Problem
You have grouped rows using groupby and want to apply a function to each group.
Solution
Combine groupby and apply:
# Load library
import pandas as pd
# Create URL
url = 'https://tinyurl.com/titanic-csv'
# Load data
dataframe = pd.read_csv(url)
# Group rows, apply function to groups
dataframe.groupby('Sex').apply(lambda x: x.count())
Name PClass Age Sex Survived SexCode
Sex
female 462
male 851
462
851
288 462 462
468 851 851
462
851
Discussion
In Recipe 3.16 I mentioned apply. apply is particularly useful when you want to
apply a function to groups. By combining groupby and apply we can calculate cus‐
tom statistics or apply any function to each group separately.
3.18 Concatenating DataFrames
Problem
You want to concatenate two DataFrames.
Solution
Use concat with axis=0 to concatenate along the row axis:
# Load library
import pandas as pd
# Create DataFrame
3.17 Applying a Function to Groups
|
55
data_a = {'id': ['1', '2', '3'],
'first': ['Alex', 'Amy', 'Allen'],
'last': ['Anderson', 'Ackerman', 'Ali']}
dataframe_a = pd.DataFrame(data_a, columns = ['id', 'first', 'last'])
# Create DataFrame
data_b = {'id': ['4', '5', '6'],
'first': ['Billy', 'Brian', 'Bran'],
'last': ['Bonder', 'Black', 'Balwner']}
dataframe_b = pd.DataFrame(data_b, columns = ['id', 'first', 'last'])
# Concatenate DataFrames by rows
pd.concat([dataframe_a, dataframe_b], axis=0)
0
1
2
0
1
2
id
1
2
3
4
5
6
first
Alex
Amy
Allen
Billy
Brian
Bran
last
Anderson
Ackerman
Ali
Bonder
Black
Balwner
You can use axis=1 to concatenate along the column axis:
# Concatenate DataFrames by columns
pd.concat([dataframe_a, dataframe_b], axis=1)
id
0 1
1 2
2 3
first
Alex
Amy
Allen
last
Anderson
Ackerman
Ali
id
4
5
6
first
Billy
Brian
Bran
last
Bonder
Black
Balwner
Discussion
Concatenating is not a word you hear much outside of computer science and pro‐
gramming, so if you have not heard it before, do not worry. The informal definition
of concatenate is to glue two objects together. In the solution we glued together two
small DataFrames using the axis parameter to indicate whether we wanted to stack
the two DataFrames on top of each other or place them side by side.
Alternatively we can use append to add a new row to a DataFrame:
# Create row
row = pd.Series([10, 'Chris', 'Chillon'], index=['id', 'first', 'last'])
# Append row
dataframe_a.append(row, ignore_index=True)
56
|
Chapter 3: Data Wrangling
0
1
2
3
id
1
2
3
10
first
Alex
Amy
Allen
Chris
last
Anderson
Ackerman
Ali
Chillon
3.19 Merging DataFrames
Problem
You want to merge two DataFrames.
Solution
To inner join, use merge with the on parameter to specify the column to merge on:
# Load library
import pandas as pd
# Create DataFrame
employee_data = {'employee_id': ['1', '2', '3', '4'],
'name': ['Amy Jones', 'Allen Keys', 'Alice Bees',
'Tim Horton']}
dataframe_employees = pd.DataFrame(employee_data, columns = ['employee_id',
'name'])
# Create DataFrame
sales_data = {'employee_id': ['3', '4', '5', '6'],
'total_sales': [23456, 2512, 2345, 1455]}
dataframe_sales = pd.DataFrame(sales_data, columns = ['employee_id',
'total_sales'])
# Merge DataFrames
pd.merge(dataframe_employees, dataframe_sales, on='employee_id')
employee_id name
total_sales
0 3
Alice Bees 23456
1 4
Tim Horton 2512
merge defaults to inner joins. If we want to do an outer join, we can specify that with
the how parameter:
3.19 Merging DataFrames
|
57
# Merge DataFrames
pd.merge(dataframe_employees, dataframe_sales, on='employee_id', how='outer')
0
1
2
3
4
5
employee_id
1
2
3
4
5
6
name
Amy Jones
Allen Keys
Alice Bees
Tim Horton
NaN
NaN
total_sales
NaN
NaN
23456.0
2512.0
2345.0
1455.0
The same parameter can be used to specify left and right joins:
# Merge DataFrames
pd.merge(dataframe_employees, dataframe_sales, on='employee_id', how='left')
0
1
2
3
employee_id
1
2
3
4
name
Amy Jones
Allen Keys
Alice Bees
Tim Horton
total_sales
NaN
NaN
23456.0
2512.0
We can also specify the column name in each DataFrame to merge on:
# Merge DataFrames
pd.merge(dataframe_employees,
dataframe_sales,
left_on='employee_id',
right_on='employee_id')
employee_id name
total_sales
0 3
Alice Bees 23456
1 4
Tim Horton 2512
If instead of merging on two columns we want to merge on the indexes of each Data‐
Frame, we can replace the left_on and right_on parameters with right_index=True
and left_index=True.
Discussion
Oftentimes, the data we need to use is complex; it doesn’t always come in one piece.
Instead in the real world, we’re usually faced with disparate datasets, from multiple
database queries or files. To get all that data into one place, we can load each data
query or data file into pandas as individual DataFrames and then merge them
together into a single DataFrame.
58
|
Chapter 3: Data Wrangling
This process might be familiar to anyone who has used SQL, a popular language for
doing merging operations (called joins). While the exact parameters used by pandas
will be different, they follow the same general patterns used by other software lan‐
guages and tools.
There are three aspects to specify with any merge operation. First, we have to specify
the two DataFrames we want to merge together. In the solution we named them data
frame_employees and dataframe_sales. Second, we have to specify the name(s) of
the columns to merge on—that is, the columns whose values are shared between the
two DataFrames. For example, in our solution both DataFrames have a column
named employee_id. To merge the two DataFrames we will match up the values in
each DataFrame’s employee_id column with each other. If these two columns use the
same name, we can use the on parameter. However, if they have different names we
can use left_on and right_on.
What is the left and right DataFrame? The simple answer is that the left DataFrame is
the first one we specified in merge and the right DataFrame is the second one. This
language comes up again in the next sets of parameters we will need.
The last aspect, and most difficult for some people to grasp, is the type of merge oper‐
ation we want to conduct. This is specified by the how parameter. merge supports the
four main types of joins:
Inner
Return only the rows that match in both DataFrames (e.g., return any row with
an employee_id value appearing in both dataframe_employees and data
frame_sales).
Outer
Return all rows in both DataFrames. If a row exists in one DataFrame but not in
the other DataFrame, fill NaN values for the missing values (e.g., return all rows
in both employee_id and dataframe_sales).
Left
Return all rows from the left DataFrame but only rows from the right DataFrame
that matched with the left DataFrame. Fill NaN values for the missing values (e.g.,
return all rows from dataframe_employees but only rows from data
frame_sales that have a value for employee_id that appears in data
frame_employees).
Right
Return all rows from the right DataFrame but only rows from the left DataFrame
that matched with the right DataFrame. Fill NaN values for the missing values
(e.g., return all rows from dataframe_sales but only rows from data
3.19 Merging DataFrames
|
59
frame_employees that have a value for employee_id that appears in data
frame_sales).
If you did not understand all of that right now, I encourage you to play around with
the how parameter in your code and see how it affects what merge returns.
See Also
• A Visual Explanation of SQL Joins
• pandas documentation on merging
60
|
Chapter 3: Data Wrangling
CHAPTER 4
Handling Numerical Data
4.0 Introduction
Quantitative data is the measurement of something—whether class size, monthly
sales, or student scores. The natural way to represent these quantities is numerically
(e.g., 29 students, $529,392 in sales). In this chapter, we will cover numerous strate‐
gies for transforming raw numerical data into features purpose-built for machine
learning algorithms.
4.1 Rescaling a Feature
Problem
You need to rescale the values of a numerical feature to be between two values.
Solution
Use scikit-learn’s MinMaxScaler to rescale a feature array:
# Load libraries
import numpy as np
from sklearn import preprocessing
# Create feature
feature = np.array([[-500.5],
[-100.1],
[0],
[100.1],
[900.9]])
# Create scaler
minmax_scale = preprocessing.MinMaxScaler(feature_range=(0, 1))
61
# Scale feature
scaled_feature = minmax_scale.fit_transform(feature)
# Show feature
scaled_feature
array([[
[
[
[
[
0.
],
0.28571429],
0.35714286],
0.42857143],
1.
]])
Discussion
Rescaling is a common preprocessing task in machine learning. Many of the algo‐
rithms described later in this book will assume all features are on the same scale, typi‐
cally 0 to 1 or –1 to 1. There are a number of rescaling techniques, but one of the
simplest is called min-max scaling. Min-max scaling uses the minimum and maxi‐
mum values of a feature to rescale values to within a range. Specifically, min-max cal‐
culates:
xi′ =
xi − min x
max x − min x
where x is the feature vector, x’i is an individual element of feature x, and x’i is the
rescaled element. In our example, we can see from the outputted array that the fea‐
ture has been successfully rescaled to between 0 and 1:
array([[ 0.
],
[ 0.28571429],
[ 0.35714286],
[ 0.42857143],
[ 1.
]])
scikit-learn’s MinMaxScaler offers two options to rescale a feature. One option is to
use fit to calculate the minimum and maximum values of the feature, then use trans
form to rescale the feature. The second option is to use fit_transform to do both
operations at once. There is no mathematical difference between the two options, but
there is sometimes a practical benefit to keeping the operations separate because it
allows us to apply the same transformation to different sets of the data.
See Also
• Feature scaling, Wikipedia
• About Feature Scaling and Normalization, Sebastian Raschka
62
|
Chapter 4: Handling Numerical Data
4.2 Standardizing a Feature
Problem
You want to transform a feature to have a mean of 0 and a standard deviation of 1.
Solution
scikit-learn’s StandardScaler performs both transformations:
# Load libraries
import numpy as np
from sklearn import preprocessing
# Create feature
x = np.array([[-1000.1],
[-200.2],
[500.5],
[600.6],
[9000.9]])
# Create scaler
scaler = preprocessing.StandardScaler()
# Transform the feature
standardized = scaler.fit_transform(x)
# Show feature
standardized
array([[-0.76058269],
[-0.54177196],
[-0.35009716],
[-0.32271504],
[ 1.97516685]])
Discussion
A common alternative to min-max scaling discussed in Recipe 4.1 is rescaling of fea‐
tures to be approximately standard normally distributed. To achieve this, we use
standardization to transform the data such that it has a mean, x̄, of 0 and a standard
deviation, σ, of 1. Specifically, each element in the feature is transformed so that:
xi′ =
xi − x
σ
4.2 Standardizing a Feature
|
63
where x’i is our standardized form of xi. The transformed feature represents the num‐
ber of standard deviations the original value is away from the feature’s mean value
(also called a z-score in statistics).
Standardization is a common go-to scaling method for machine learning preprocess‐
ing and in my experience is used more than min-max scaling. However, it depends on
the learning algorithm. For example, principal component analysis often works better
using standardization, while min-max scaling is often recommended for neural net‐
works (both algorithms are discussed later in this book). As a general rule, I’d recom‐
mend defaulting to standardization unless you have a specific reason to use an alter‐
native.
We can see the effect of standardization by looking at the mean and standard devia‐
tion of our solution’s output:
# Print mean and standard deviation
print("Mean:", round(standardized.mean()))
print("Standard deviation:", standardized.std())
Mean: 0.0
Standard deviation: 1.0
If our data has significant outliers, it can negatively impact our standardization by
affecting the feature’s mean and variance. In this scenario, it is often helpful to instead
rescale the feature using the median and quartile range. In scikit-learn, we do this
using the RobustScaler method:
# Create scaler
robust_scaler = preprocessing.RobustScaler()
# Transform feature
robust_scaler.fit_transform(x)
array([[ -1.87387612],
[ -0.875
],
[ 0.
],
[ 0.125
],
[ 10.61488511]])
4.3 Normalizing Observations
Problem
You want to rescale the feature values of observations to have unit norm (a total
length of 1).
Solution
Use Normalizer with a norm argument:
64
|
Chapter 4: Handling Numerical Data
# Load libraries
import numpy as np
from sklearn.preprocessing import Normalizer
# Create feature matrix
features = np.array([[0.5, 0.5],
[1.1, 3.4],
[1.5, 20.2],
[1.63, 34.4],
[10.9, 3.3]])
# Create normalizer
normalizer = Normalizer(norm="l2")
# Transform feature matrix
normalizer.transform(features)
array([[
[
[
[
[
0.70710678,
0.30782029,
0.07405353,
0.04733062,
0.95709822,
0.70710678],
0.95144452],
0.99725427],
0.99887928],
0.28976368]])
Discussion
Many rescaling methods (e.g., min-max scaling and standardization) operate on fea‐
tures; however, we can also rescale across individual observations. Normalizer
rescales the values on individual observations to have unit norm (the sum of their
lengths is 1). This type of rescaling is often used when we have many equivalent fea‐
tures (e.g., text classification when every word or n-word group is a feature).
Normalizer provides three norm options with Euclidean norm (often called L2)
being the default argument:
∥ x ∥2 = x21 + x22 + ⋯ + x2n
where x is an individual observation and xn is that observation’s value for the nth fea‐
ture.
# Transform feature matrix
features_l2_norm = Normalizer(norm="l2").transform(features)
# Show feature matrix
features_l2_norm
array([[
[
[
[
[
0.70710678,
0.30782029,
0.07405353,
0.04733062,
0.95709822,
0.70710678],
0.95144452],
0.99725427],
0.99887928],
0.28976368]])
4.3 Normalizing Observations
|
65
Alternatively, we can specify Manhattan norm (L1):
∥ x ∥1 =
n
∑ xi
i=1
.
# Transform feature matrix
features_l1_norm = Normalizer(norm="l1").transform(features)
# Show feature matrix
features_l1_norm
array([[
[
[
[
[
0.5
,
0.24444444,
0.06912442,
0.04524008,
0.76760563,
0.5
],
0.75555556],
0.93087558],
0.95475992],
0.23239437]])
Intuitively, L2 norm can be thought of as the distance between two points in New
York for a bird (i.e., a straight line), while L1 can be thought of as the distance for a
human walking on the street (walk north one block, east one block, north one block,
east one block, etc.), which is why it is called “Manhattan norm” or “Taxicab norm.”
Practically, notice that norm='l1' rescales an observation’s values so they sum to 1,
which can sometimes be a desirable quality:
# Print sum
print("Sum of the first observation\'s values:",
features_l1_norm[0, 0] + features_l1_norm[0, 1])
Sum of the first observation's values: 1.0
4.4 Generating Polynomial and Interaction Features
Problem
You want to create polynominal and interaction features.
Solution
Even though some choose to create polynomial and interaction features manually,
scikit-learn offers a built-in method:
# Load libraries
import numpy as np
from sklearn.preprocessing import PolynomialFeatures
# Create feature matrix
features = np.array([[2, 3],
[2, 3],
66
|
Chapter 4: Handling Numerical Data
[2, 3]])
# Create PolynomialFeatures object
polynomial_interaction = PolynomialFeatures(degree=2, include_bias=False)
# Create polynomial features
polynomial_interaction.fit_transform(features)
array([[ 2.,
[ 2.,
[ 2.,
3.,
3.,
3.,
4.,
4.,
4.,
6.,
6.,
6.,
9.],
9.],
9.]])
The degree parameter determines the maximum degree of the polynomial. For
example, degree=2 will create new features raised to the second power:
x1, x2, x21, x22
while degree=3 will create new features raised to the second and third power:
x1, x2, x21, x22, x31, x32
Furthermore, by default PolynomialFeatures includes interaction features:
x1x2
We can restrict the features created to only interaction features by setting interac
tion_only to True:
interaction = PolynomialFeatures(degree=2,
interaction_only=True, include_bias=False)
interaction.fit_transform(features)
array([[ 2.,
[ 2.,
[ 2.,
3.,
3.,
3.,
6.],
6.],
6.]])
Discussion
Polynomial features are often created when we want to include the notion that there
exists a nonlinear relationship between the features and the target. For example, we
might suspect that the effect of age on the probability of having a major medical con‐
dition is not constant over time but increases as age increases. We can encode that
nonconstant effect in a feature, x, by generating that feature’s higher-order forms (x2,
x3, etc.).
4.4 Generating Polynomial and Interaction Features
|
67
Additionally, often we run into situations where the effect of one feature is dependent
on another feature. A simple example would be if we were trying to predict whether
or not our coffee was sweet and we had two features: 1) whether or not the coffee was
stirred and 2) if we added sugar. Individually, each feature does not predict coffee
sweetness, but the combination of their effects does. That is, a coffee would only be
sweet if the coffee had sugar and was stirred. The effects of each feature on the target
(sweetness) are dependent on each other. We can encode that relationship by includ‐
ing an interaction feature that is the product of the individual features.
4.5 Transforming Features
Problem
You want to make a custom transformation to one or more features.
Solution
In scikit-learn, use FunctionTransformer to apply a function to a set of features:
# Load libraries
import numpy as np
from sklearn.preprocessing import FunctionTransformer
# Create feature matrix
features = np.array([[2, 3],
[2, 3],
[2, 3]])
# Define a simple function
def add_ten(x):
return x + 10
# Create transformer
ten_transformer = FunctionTransformer(add_ten)
# Transform feature matrix
ten_transformer.transform(features)
array([[12, 13],
[12, 13],
[12, 13]])
We can create the same transformation in pandas using apply:
# Load library
import pandas as pd
# Create DataFrame
df = pd.DataFrame(features, columns=["feature_1", "feature_2"])
68
|
Chapter 4: Handling Numerical Data
# Apply function
df.apply(add_ten)
feature_1
0 12
1 12
2 12
feature_2
13
13
13
Discussion
It is common to want to make some custom transformations to one or more features.
For example, we might want to create a feature that is the natural log of the values of
the different feature. We can do this by creating a function and then mapping it to
features using either scikit-learn’s FunctionTransformer or pandas’ apply. In the sol‐
ution we created a very simple function, add_ten, which added 10 to each input, but
there is no reason we could not define a much more complex function.
4.6 Detecting Outliers
Problem
You want to identify extreme observations.
Solution
Detecting outliers is unfortunately more of an art than a science. However, a common
method is to assume the data is normally distributed and based on that assumption
“draw” an ellipse around the data, classifying any observation inside the ellipse as an
inlier (labeled as 1) and any observation outside the ellipse as an outlier (labeled as
-1):
# Load libraries
import numpy as np
from sklearn.covariance import EllipticEnvelope
from sklearn.datasets import make_blobs
# Create simulated data
features, _ = make_blobs(n_samples = 10,
n_features = 2,
centers = 1,
random_state = 1)
# Replace the first observation's values with extreme values
features[0,0] = 10000
features[0,1] = 10000
# Create detector
4.6 Detecting Outliers
|
69
outlier_detector = EllipticEnvelope(contamination=.1)
# Fit detector
outlier_detector.fit(features)
# Predict outliers
outlier_detector.predict(features)
array([-1,
1,
1,
1,
1,
1,
1,
1,
1,
1])
A major limitation of this approach is the need to specify a contamination parame‐
ter, which is the proportion of observations that are outliers—a value that we don’t
know. Think of contamination as our estimate of the cleanliness of our data. If we
expect our data to have few outliers, we can set contamination to something small.
However, if we believe that the data is very likely to have outliers, we can set it to a
higher value.
Instead of looking at observations as a whole, we can instead look at individual fea‐
tures and identify extreme values in those features using interquartile range (IQR):
# Create one feature
feature = features[:,0]
# Create a function to return index of outliers
def indicies_of_outliers(x):
q1, q3 = np.percentile(x, [25, 75])
iqr = q3 - q1
lower_bound = q1 - (iqr * 1.5)
upper_bound = q3 + (iqr * 1.5)
return np.where((x > upper_bound) | (x < lower_bound))
# Run function
indicies_of_outliers(feature)
(array([0]),)
IQR is the difference between the first and third quartile of a set of data. You can
think of IQR as the spread of the bulk of the data, with outliers being observations far
from the main concentration of data. Outliers are commonly defined as any value 1.5
IQRs less than the first quartile or 1.5 IQRs greater than the third quartile.
Discussion
There is no single best technique for detecting outliers. Instead, we have a collection
of techniques all with their own advantages and disadvantages. Our best strategy is
often trying multiple techniques (e.g., both EllipticEnvelope and IQR-based detec‐
tion) and looking at the results as a whole.
If at all possible, we should take a look at observations we detect as outliers and try to
understand them. For example, if we have a dataset of houses and one feature is num‐
70
|
Chapter 4: Handling Numerical Data
ber of rooms, is an outlier with 100 rooms really a house or is it actually a hotel that
has been misclassified?
See Also
• Three ways to detect outliers (and the source of the IQR function used in this
recipe)
4.7 Handling Outliers
Problem
You have outliers.
Solution
Typically we have three strategies we can use to handle outliers. First, we can drop
them:
# Load library
import pandas as pd
# Create DataFrame
houses = pd.DataFrame()
houses['Price'] = [534433, 392333, 293222, 4322032]
houses['Bathrooms'] = [2, 3.5, 2, 116]
houses['Square_Feet'] = [1500, 2500, 1500, 48000]
# Filter observations
houses[houses['Bathrooms'] < 20]
Price Bathrooms
0 534433 2.0
1 392333 3.5
2 293222 2.0
Square_Feet
1500
2500
1500
Second, we can mark them as outliers and include it as a feature:
# Load library
import numpy as np
# Create feature based on boolean condition
houses["Outlier"] = np.where(houses["Bathrooms"] < 20, 0, 1)
# Show data
houses
4.7 Handling Outliers
|
71
0
1
2
3
Price
534433
392333
293222
4322032
Bathrooms
2.0
3.5
2.0
116.0
Square_Feet
1500
2500
1500
48000
Outlier
0
0
0
1
Finally, we can transform the feature to dampen the effect of the outlier:
# Log feature
houses["Log_Of_Square_Feet"] = [np.log(x) for x in houses["Square_Feet"]]
# Show data
houses
0
1
2
3
Price
534433
392333
293222
4322032
Bathrooms
2.0
3.5
2.0
116.0
Square_Feet
1500
2500
1500
48000
Outlier
0
0
0
1
Log_Of_Square_Feet
7.313220
7.824046
7.313220
10.778956
Discussion
Similar to detecting outliers, there is no hard-and-fast rule for handling them. How
we handle them should be based on two aspects. First, we should consider what
makes them an outlier. If we believe they are errors in the data such as from a broken
sensor or a miscoded value, then we might drop the observation or replace outlier
values with NaN since we can’t believe those values. However, if we believe the outliers
are genuine extreme values (e.g., a house [mansion] with 200 bathrooms), then mark‐
ing them as outliers or transforming their values is more appropriate.
Second, how we handle outliers should be based on our goal for machine learning.
For example, if we want to predict house prices based on features of the house, we
might reasonably assume the price for mansions with over 100 bathrooms is driven
by a different dynamic than regular family homes. Furthermore, if we are training a
model to use as part of an online home loan web application, we might assume that
our potential users will not include billionaires looking to buy a mansion.
So what should we do if we have outliers? Think about why they are outliers, have an
end goal in mind for the data, and, most importantly, remember that not making a
decision to address outliers is itself a decision with implications.
One additional point: if you do have outliers standardization might not be appropri‐
ate because the mean and variance might be highly influenced by the outliers. In this
case, use a rescaling method more robust against outliers like RobustScaler.
72
|
Chapter 4: Handling Numerical Data
See Also
• RobustScaler documentation
4.8 Discretizating Features
Problem
You have a numerical feature and want to break it up into discrete bins.
Solution
Depending on how we want to break up the data, there are two techniques we can
use. First, we can binarize the feature according to some threshold:
# Load libraries
import numpy as np
from sklearn.preprocessing import Binarizer
# Create feature
age = np.array([[6],
[12],
[20],
[36],
[65]])
# Create binarizer
binarizer = Binarizer(18)
# Transform feature
binarizer.fit_transform(age)
array([[0],
[0],
[1],
[1],
[1]])
Second, we can break up numerical features according to multiple thresholds:
# Bin feature
np.digitize(age, bins=[20,30,64])
array([[0],
[0],
[1],
[2],
[3]])
4.8 Discretizating Features
|
73
Note that the arguments for the bins parameter denote the left edge of each bin. For
example, the 20 argument does not include the element with the value of 20, only the
two values smaller than 20. We can switch this behavior by setting the parameter
right to True:
# Bin feature
np.digitize(age, bins=[20,30,64], right=True)
array([[0],
[0],
[0],
[2],
[3]])
Discussion
Discretization can be a fruitful strategy when we have reason to believe that a numeri‐
cal feature should behave more like a categorical feature. For example, we might
believe there is very little difference in the spending habits of 19- and 20-year-olds,
but a significant difference between 20- and 21-year-olds (the age in the United States
when young adults can consume alcohol). In that example, it could be useful to break
up individuals in our data into those who can drink alcohol and those who cannot.
Similarly, in other cases it might be useful to discretize our data into three or more
bins.
In the solution, we saw two methods of discretization—scikit-learn’s Binarizer for
two bins and NumPy’s digitize for three or more bins—however, we can also use
digitize to binarize features like Binarizer by only specifying a single threshold:
# Bin feature
np.digitize(age, bins=[18])
array([[0],
[0],
[1],
[1],
[1]])
See Also
• digitize documentation
4.9 Grouping Observations Using Clustering
Problem
You want to cluster observations so that similar observations are grouped together.
74
|
Chapter 4: Handling Numerical Data
Solution
If you know that you have k groups, you can use k-means clustering to group similar
observations and output a new feature containing each observation’s group member‐
ship:
# Load libraries
import pandas as pd
from sklearn.datasets import make_blobs
from sklearn.cluster import KMeans
# Make simulated feature matrix
features, _ = make_blobs(n_samples = 50,
n_features = 2,
centers = 3,
random_state = 1)
# Create DataFrame
dataframe = pd.DataFrame(features, columns=["feature_1", "feature_2"])
# Make k-means clusterer
clusterer = KMeans(3, random_state=0)
# Fit clusterer
clusterer.fit(features)
# Predict values
dataframe["group"] = clusterer.predict(features)
# View first few observations
dataframe.head(5)
0
1
2
3
4
feature_1
–9.877554
–7.287210
–6.943061
–7.440167
–6.641388
feature_2
–3.336145
–8.353986
–7.023744
–8.791959
–8.075888
group
0
2
2
2
2
Discussion
We are jumping ahead of ourselves a bit and will go much more in depth about clus‐
tering algorithms later in the book. However, I wanted to point out that we can use
clustering as a preprocessing step. Specifically, we use unsupervised learning algo‐
rithms like k-means to cluster observations into groups. The end result is a categori‐
cal feature with similar observations being members of the same group.
4.9 Grouping Observations Using Clustering
|
75
Don’t worry if you did not understand all of that right now: just file away the idea that
clustering can be used in preprocessing. And if you really can’t wait, feel free to flip to
Chapter 19 now.
4.10 Deleting Observations with Missing Values
Problem
You need to delete observations containing missing values.
Solution
Deleting observations with missing values is easy with a clever line of NumPy:
# Load library
import numpy as np
# Create feature matrix
features = np.array([[1.1, 11.1],
[2.2, 22.2],
[3.3, 33.3],
[4.4, 44.4],
[np.nan, 55]])
# Keep only observations that are not (denoted by ~) missing
features[~np.isnan(features).any(axis=1)]
array([[
[
[
[
1.1,
2.2,
3.3,
4.4,
11.1],
22.2],
33.3],
44.4]])
Alternatively, we can drop missing observations using pandas:
# Load library
import pandas as pd
# Load data
dataframe = pd.DataFrame(features, columns=["feature_1", "feature_2"])
# Remove observations with missing values
dataframe.dropna()
0
1
2
3
76
feature_1
1.1
2.2
3.3
4.4
|
feature_2
11.1
22.2
33.3
44.4
Chapter 4: Handling Numerical Data
Discussion
Most machine learning algorithms cannot handle any missing values in the target and
feature arrays. For this reason, we cannot ignore missing values in our data and must
address the issue during preprocessing.
The simplest solution is to delete every observation that contains one or more miss‐
ing values, a task quickly and easily accomplished using NumPy or pandas.
That said, we should be very reluctant to delete observations with missing values.
Deleting them is the nuclear option, since our algorithm loses access to the informa‐
tion contained in the observation’s non-missing values.
Just as important, depending on the cause of the missing values, deleting observations
can introduce bias into our data. There are three types of missing data:
Missing Completely At Random (MCAR)
The probability that a value is missing is independent of everything. For example,
a survey respondent rolls a die before answering a question: if she rolls a six, she
skips that question.
Missing At Random (MAR)
The probability that a value is missing is not completely random, but depends on
the information captured in other features. For example, a survey asks about gen‐
der identity and annual salary and women are more likely to skip the salary ques‐
tion; however, their nonresponse depends only on information we have captured
in our gender identity feature.
Missing Not At Random (MNAR)
The probability that a value is missing is not random and depends on informa‐
tion not captured in our features. For example, a survey asks about gender iden‐
tity and women are more likely to skip the salary question, and we do not have a
gender identity feature in our data.
It is sometimes acceptable to delete observations if they are MCAR or MAR. How‐
ever, if the value is MNAR, the fact that a value is missing is itself information. Delet‐
ing MNAR observations can inject bias into our data because we are removing obser‐
vations produced by some unobserved systematic effect.
See Also
• Identifying the Three Types of Missing Data
• Missing-Data Imputation
4.10 Deleting Observations with Missing Values
|
77
4.11 Imputing Missing Values
Problem
You have missing values in your data and want to fill in or predict their values.
Solution
If you have a small amount of data, predict the missing values using k-nearest neigh‐
bors (KNN):
# Load libraries
import numpy as np
from fancyimpute import KNN
from sklearn.preprocessing import StandardScaler
from sklearn.datasets import make_blobs
# Make a simulated feature matrix
features, _ = make_blobs(n_samples = 1000,
n_features = 2,
random_state = 1)
# Standardize the features
scaler = StandardScaler()
standardized_features = scaler.fit_transform(features)
# Replace the first feature's first value with a missing value
true_value = standardized_features[0,0]
standardized_features[0,0] = np.nan
# Predict the missing values in the feature matrix
features_knn_imputed = KNN(k=5, verbose=0).complete(standardized_features)
# Compare true and imputed values
print("True Value:", true_value)
print("Imputed Value:", features_knn_imputed[0,0])
True Value: 0.8730186114
Imputed Value: 1.09553327131
Alternatively, we can use scikit-learn’s Imputer module to fill in missing values with
the feature’s mean, median, or most frequent value. However, we will typically get
worse results than KNN:
# Load library
from sklearn.preprocessing import Imputer
# Create imputer
mean_imputer = Imputer(strategy="mean", axis=0)
# Impute values
78
|
Chapter 4: Handling Numerical Data
features_mean_imputed = mean_imputer.fit_transform(features)
# Compare true and imputed values
print("True Value:", true_value)
print("Imputed Value:", features_mean_imputed[0,0])
True Value: 0.8730186114
Imputed Value: -3.05837272461
Discussion
There are two main strategies for replacing missing data with substitute values, each
of which has strengths and weaknesses. First, we can use machine learning to predict
the values of the missing data. To do this we treat the feature with missing values as a
target vector and use the remaining subset of features to predict missing values.
While we can use a wide range of machine learning algorithms to impute values, a
popular choice is KNN. KNN is addressed in depth later in Chapter 14, but the short
explanation is that the algorithm uses the k nearest observations (according to some
distance metric) to predict the missing value. In our solution we predicted the miss‐
ing value using the five closest observations.
The downside to KNN is that in order to know which observations are the closest to
the missing value, it needs to calculate the distance between the missing value and
every single observation. This is reasonable in smaller datasets, but quickly becomes
problematic if a dataset has millions of observations.
An alternative and more scalable strategy is to fill in all missing values with some
average value. For example, in our solution we used scikit-learn to fill in missing val‐
ues with a feature’s mean value. The imputed value is often not as close to the true
value as when we used KNN, but we can scale mean-filling to data containing mil‐
lions of observations easily.
If we use imputation, it is a good idea to create a binary feature indicating whether or
not the observation contains an imputed value.
See Also
• A Study of K-Nearest Neighbour as an Imputation Method
4.11 Imputing Missing Values
|
79
CHAPTER 5
Handling Categorical Data
5.0 Introduction
It is often useful to measure objects not in terms of their quantity but in terms of
some quality. We frequently represent this qualitative information as an observation’s
membership in a discrete category such as gender, colors, or brand of car. However,
not all categorical data is the same. Sets of categories with no intrinsic ordering is
called nominal. Examples of nominal categories include:
• Blue, Red, Green
• Man, Woman
• Banana, Strawberry, Apple
In contrast, when a set of categories has some natural ordering we refer to it as ordi‐
nal. For example:
• Low, Medium, High
• Young, Old
• Agree, Neutral, Disagree
Furthermore, categorical information is often represented in data as a vector or col‐
umn of strings (e.g., "Maine", "Texas", "Delaware"). The problem is that most
machine learning algorithms require inputs be numerical values.
The k-nearest neighbor algorithm provides a simple example. One step in the algo‐
rithm is calculating the distances between observations—often using Euclidean
distance:
81
∑ni = 1
xi − yi
2
where x and y are two observations and subscript i denotes the value for the observa‐
tions’ ith feature. However, the distance calculation obviously is impossible if the
value of xi is a string (e.g., "Texas"). Instead, we need to convert the string into some
numerical format so that it can be inputted into the Euclidean distance equation. Our
goal is to make a transformation that properly conveys the information in the cate‐
gories (ordinality, relative intervals between categories, etc.). In this chapter we will
cover techniques for making this transformation as well as overcoming other chal‐
lenges often encountered when handling categorical data.
5.1 Encoding Nominal Categorical Features
Problem
You have a feature with nominal classes that has no intrinsic ordering (e.g., apple,
pear, banana).
Solution
One-hot encode the feature using scikit-learn’s LabelBinarizer:
# Import libraries
import numpy as np
from sklearn.preprocessing import LabelBinarizer, MultiLabelBinarizer
# Create feature
feature = np.array([["Texas"],
["California"],
["Texas"],
["Delaware"],
["Texas"]])
# Create one-hot encoder
one_hot = LabelBinarizer()
# One-hot encode feature
one_hot.fit_transform(feature)
array([[0,
[1,
[0,
[0,
[0,
0,
0,
0,
1,
0,
1],
0],
1],
0],
1]])
We can use the classes_ method to output the classes:
82
|
Chapter 5: Handling Categorical Data
# View feature classes
one_hot.classes_
array(['California', 'Delaware', 'Texas'],
dtype='<U10')
If we want to reverse the one-hot encoding, we can use inverse_transform:
# Reverse one-hot encoding
one_hot.inverse_transform(one_hot.transform(feature))
array(['Texas', 'California', 'Texas', 'Delaware', 'Texas'],
dtype='<U10')
We can even use pandas to one-hot encode the feature:
# Import library
import pandas as pd
# Create dummy variables from feature
pd.get_dummies(feature[:,0])
0
1
2
3
4
California
0
1
0
0
0
Delaware
0
0
0
1
0
Texas
1
0
1
0
1
One helpful ability of scikit-learn is to handle a situation where each observation lists
multiple classes:
# Create multiclass feature
multiclass_feature = [("Texas", "Florida"),
("California", "Alabama"),
("Texas", "Florida"),
("Delware", "Florida"),
("Texas", "Alabama")]
# Create multiclass one-hot encoder
one_hot_multiclass = MultiLabelBinarizer()
# One-hot encode multiclass feature
one_hot_multiclass.fit_transform(multiclass_feature)
array([[0,
[1,
[0,
[0,
[1,
0,
1,
0,
0,
0,
0,
0,
0,
1,
0,
1,
0,
1,
1,
0,
1],
0],
1],
0],
1]])
Once again, we can see the classes with the classes_ method:
5.1 Encoding Nominal Categorical Features
|
83
# View classes
one_hot_multiclass.classes_
array(['Alabama', 'California', 'Delware', 'Florida', 'Texas'], dtype=object)
Discussion
We might think the proper strategy is to assign each class a numerical value (e.g.,
Texas = 1, California = 2). However, when our classes have no intrinsic ordering (e.g.,
Texas isn’t “less” than California), our numerical values erroneously create an order‐
ing that is not present.
The proper strategy is to create a binary feature for each class in the original feature.
This is often called one-hot encoding (in machine learning literature) or dummying (in
statistical and research literature). Our solution’s feature was a vector containing three
classes (i.e., Texas, California, and Delaware). In one-hot encoding, each class
becomes its own feature with 1s when the class appears and 0s otherwise. Because our
feature had three classes, one-hot encoding returned three binary features (one for
each class). By using one-hot encoding we can capture the membership of an obser‐
vation in a class while preserving the notion that the class lacks any sort of hierarchy.
Finally, it is worthwhile to note that it is often recommended that after one-hot
encoding a feature, we drop one of the one-hot encoded features in the resulting
matrix to avoid linear dependence.
See Also
• Dummy Variable Trap, Algosome
• Dropping one of the columns when using one-hot encoding, CrossValidated
5.2 Encoding Ordinal Categorical Features
Problem
You have an ordinal categorical feature (e.g., high, medium, low).
Solution
Use pandas DataFrame’s replace method to transform string labels to numerical
equivalents:
# Load library
import pandas as pd
# Create features
dataframe = pd.DataFrame({"Score": ["Low", "Low", "Medium", "Medium", "High"]})
84
|
Chapter 5: Handling Categorical Data
# Create mapper
scale_mapper = {"Low":1,
"Medium":2,
"High":3}
# Replace feature values with scale
dataframe["Score"].replace(scale_mapper)
0
1
1
1
2
2
3
2
4
3
Name: Score, dtype: int64
Discussion
Often we have a feature with classes that have some kind of natural ordering. A
famous example is the Likert scale:
• Strongly Agree
• Agree
• Neutral
• Disagree
• Strongly Disagree
When encoding the feature for use in machine learning, we need to transform the
ordinal classes into numerical values that maintain the notion of ordering. The most
common approach is to create a dictionary that maps the string label of the class to a
number and then apply that map to the feature.
It is important that our choice of numeric values is based on our prior information on
the ordinal classes. In our solution, high is literally three times larger than low. This
is fine in any instances, but can break down if the assumed intervals between the
classes are not equal:
dataframe = pd.DataFrame({"Score": ["Low",
"Low",
"Medium",
"Medium",
"High",
"Barely More Than Medium"]})
scale_mapper = {"Low":1,
"Medium":2,
"Barely More Than Medium": 3,
"High":4}
5.2 Encoding Ordinal Categorical Features
|
85
dataframe["Score"].replace(scale_mapper)
0
1
1
1
2
2
3
2
4
4
5
3
Name: Score, dtype: int64
In this example, the distance between Low and Medium is the same as the distance
between Medium and Barely More Than Medium, which is almost certainly not accu‐
rate. The best approach is to be conscious about the numerical values mapped to
classes:
scale_mapper = {"Low":1,
"Medium":2,
"Barely More Than Medium": 2.1,
"High":3}
dataframe["Score"].replace(scale_mapper)
0
1.0
1
1.0
2
2.0
3
2.0
4
3.0
5
2.1
Name: Score, dtype: float64
5.3 Encoding Dictionaries of Features
Problem
You have a dictionary and want to convert it into a feature matrix.
Solution
Use DictVectorizer:
# Import library
from sklearn.feature_extraction import DictVectorizer
# Create dictionary
data_dict = [{"Red":
{"Red":
{"Red":
{"Red":
2,
4,
1,
2,
"Blue": 4},
"Blue": 3},
"Yellow": 2},
"Yellow": 2}]
# Create dictionary vectorizer
86
|
Chapter 5: Handling Categorical Data
dictvectorizer = DictVectorizer(sparse=False)
# Convert dictionary to feature matrix
features = dictvectorizer.fit_transform(data_dict)
# View feature matrix
features
array([[
[
[
[
4.,
3.,
0.,
0.,
2.,
4.,
1.,
2.,
0.],
0.],
2.],
2.]])
By default DictVectorizer outputs a sparse matrix that only stores elements with a
value other than 0. This can be very helpful when we have massive matrices (often
encountered in natural language processing) and want to minimize the memory
requirements. We can force DictVectorizer to output a dense matrix using
sparse=False.
We can get the names of each generated feature using the get_feature_names
method:
# Get feature names
feature_names = dictvectorizer.get_feature_names()
# View feature names
feature_names
['Blue', 'Red', 'Yellow']
While not necessary, for the sake of illustration we can create a pandas DataFrame to
view the output better:
# Import library
import pandas as pd
# Create dataframe from features
pd.DataFrame(features, columns=feature_names)
0
1
2
3
Blue
4.0
3.0
0.0
0.0
Red
2.0
4.0
1.0
2.0
Yellow
0.0
0.0
2.0
2.0
Discussion
A dictionary is a popular data structure used by many programming languages; how‐
ever, machine learning algorithms expect the data to be in the form of a matrix. We
can accomplish this using scikit-learn’s dictvectorizer.
5.3 Encoding Dictionaries of Features
|
87
This is a common situation when working with natural language processing. For
example, we might have a collection of documents and for each document we have a
dictionary containing the number of times every word appears in the document.
Using dictvectorizer, we can easily create a feature matrix where every feature is
the number of times a word appears in each document:
# Create word counts dictionaries for four documents
doc_1_word_count = {"Red": 2, "Blue": 4}
doc_2_word_count = {"Red": 4, "Blue": 3}
doc_3_word_count = {"Red": 1, "Yellow": 2}
doc_4_word_count = {"Red": 2, "Yellow": 2}
# Create list
doc_word_counts = [doc_1_word_count,
doc_2_word_count,
doc_3_word_count,
doc_4_word_count]
# Convert list of word count dictionaries into feature matrix
dictvectorizer.fit_transform(doc_word_counts)
array([[
[
[
[
4.,
3.,
0.,
0.,
2.,
4.,
1.,
2.,
0.],
0.],
2.],
2.]])
In our toy example there are only three unique words (Red, Yellow, Blue) so there are
only three features in our matrix; however, you can imagine that if each document
was actually a book in a university library our feature matrix would be very large (and
then we would want to set spare to True).
See Also
• How to use dictionaries in Python
• SciPy Sparse Matrices
5.4 Imputing Missing Class Values
Problem
You have a categorical feature containing missing values that you want to replace with
predicted values.
88
|
Chapter 5: Handling Categorical Data
Solution
The ideal solution is to train a machine learning classifier algorithm to predict the
missing values, commonly a k-nearest neighbors (KNN) classifier:
# Load libraries
import numpy as np
from sklearn.neighbors import KNeighborsClassifier
# Create feature matrix with categorical feature
X = np.array([[0, 2.10, 1.45],
[1, 1.18, 1.33],
[0, 1.22, 1.27],
[1, -0.21, -1.19]])
# Create feature matrix with missing values in the categorical feature
X_with_nan = np.array([[np.nan, 0.87, 1.31],
[np.nan, -0.67, -0.22]])
# Train KNN learner
clf = KNeighborsClassifier(3, weights='distance')
trained_model = clf.fit(X[:,1:], X[:,0])
# Predict missing values' class
imputed_values = trained_model.predict(X_with_nan[:,1:])
# Join column of predicted class with their other features
X_with_imputed = np.hstack((imputed_values.reshape(-1,1), X_with_nan[:,1:]))
# Join two feature matrices
np.vstack((X_with_imputed, X))
array([[
[
[
[
[
[
0.
1.
0.
1.
0.
1.
, 0.87, 1.31],
, -0.67, -0.22],
, 2.1 , 1.45],
, 1.18, 1.33],
, 1.22, 1.27],
, -0.21, -1.19]])
An alternative solution is to fill in missing values with the feature’s most frequent
value:
from sklearn.preprocessing import Imputer
# Join the two feature matrices
X_complete = np.vstack((X_with_nan, X))
imputer = Imputer(strategy='most_frequent', axis=0)
imputer.fit_transform(X_complete)
array([[ 0.
[ 0.
, 0.87, 1.31],
, -0.67, -0.22],
5.4 Imputing Missing Class Values
|
89
[
[
[
[
0.
1.
0.
1.
, 2.1 , 1.45],
, 1.18, 1.33],
, 1.22, 1.27],
, -0.21, -1.19]])
Discussion
When we have missing values in a categorical feature, our best solution is to open our
toolbox of machine learning algorithms to predict the values of the missing observa‐
tions. We can accomplish this by treating the feature with the missing values as the
target vector and the other features as the feature matrix. A commonly used algo‐
rithm is KNN (discussed in depth later in this book), which assigns to the missing
value the median class of the k nearest observations.
Alternatively, we can fill in missing values with the most frequent class of the feature.
While less sophisticated than KNN, it is much more scalable to larger data. In either
case, it is advisable to include a binary feature indicating which observations contain
imputed values.
See Also
• Overcoming Missing Values in a Random Forest Classifier
• A Study of K-Nearest Neighbour as an Imputation Method
5.5 Handling Imbalanced Classes
Problem
You have a target vector with highly imbalanced classes.
Solution
Collect more data. If that isn’t possible, change the metrics used to evaluate your
model. If that doesn’t work, consider using a model’s built-in class weight parameters
(if available), downsampling, or upsampling. We cover evaluation metrics in a later
chapter, so for now let us focus on class weight parameters, downsampling, and
upsampling.
To demonstrate our solutions, we need to create some data with imbalanced classes.
Fisher’s Iris dataset contains three balanced classes of 50 observations, each indicating
the species of flower (Iris setosa, Iris virginica, and Iris versicolor). To unbalance the
dataset, we remove 40 of the 50 Iris setosa observations and then merge the Iris virgin‐
ica and Iris versicolor classes. The end result is a binary target vector indicating if an
90
|
Chapter 5: Handling Categorical Data
observation is an Iris setosa flower or not. The result is 10 observations of Iris setosa
(class 0) and 100 observations of not Iris setosa (class 1):
# Load libraries
import numpy as np
from sklearn.ensemble import RandomForestClassifier
from sklearn.datasets import load_iris
# Load iris data
iris = load_iris()
# Create feature matrix
features = iris.data
# Create target vector
target = iris.target
# Remove first 40 observations
features = features[40:,:]
target = target[40:]
# Create binary target vector indicating if class 0
target = np.where((target == 0), 0, 1)
# Look at the imbalanced target vector
target
array([0,
1,
1,
1,
1,
0,
1,
1,
1,
1,
0,
1,
1,
1,
1,
0,
1,
1,
1,
1,
0,
1,
1,
1,
1,
0,
1,
1,
1,
1,
0,
1,
1,
1,
1,
0,
1,
1,
1,
1,
0,
1,
1,
1,
1,
0,
1,
1,
1,
1,
1,
1,
1,
1,
1,
1,
1,
1,
1,
1,
1,
1,
1,
1,
1,
1,
1,
1,
1,
1,
1,
1,
1,
1,
1,
1,
1,
1,
1,
1,
1,
1,
1,
1,
1,
1, 1,
1, 1,
1, 1,
1, 1,
1])
1,
1,
1,
1,
1,
1,
1,
1,
1,
1,
1,
1,
1,
1,
1,
1,
Many algorithms in scikit-learn offer a parameter to weight classes during training to
counteract the effect of their imbalance. While we have not covered it yet, RandomFor
estClassifier is a popular classification algorithm and includes a class_weight
parameter. You can pass an argument specifying the desired class weights explicitly:
# Create weights
weights = {0: .9, 1: 0.1}
# Create random forest classifier with weights
RandomForestClassifier(class_weight=weights)
RandomForestClassifier(bootstrap=True, class_weight={0: 0.9, 1: 0.1},
criterion='gini', max_depth=None, max_features='auto',
max_leaf_nodes=None, min_impurity_decrease=0.0,
min_impurity_split=None, min_samples_leaf=1,
min_samples_split=2, min_weight_fraction_leaf=0.0,
n_estimators=10, n_jobs=1, oob_score=False, random_state=None,
verbose=0, warm_start=False)
5.5 Handling Imbalanced Classes
|
91
Or you can pass balanced, which automatically creates weights inversely propor‐
tional to class frequencies:
# Train a random forest with balanced class weights
RandomForestClassifier(class_weight="balanced")
RandomForestClassifier(bootstrap=True, class_weight='balanced',
criterion='gini', max_depth=None, max_features='auto',
max_leaf_nodes=None, min_impurity_decrease=0.0,
min_impurity_split=None, min_samples_leaf=1,
min_samples_split=2, min_weight_fraction_leaf=0.0,
n_estimators=10, n_jobs=1, oob_score=False, random_state=None,
verbose=0, warm_start=False)
Alternatively, we can downsample the majority class or upsample the minority class.
In downsampling, we randomly sample without replacement from the majority class
(i.e., the class with more observations) to create a new subset of observations equal in
size to the minority class. For example, if the minority class has 10 observations, we
will randomly select 10 observations from the majority class and use those 20 obser‐
vations as our data. Here we do exactly that using our unbalanced Iris data:
# Indicies of each class' observations
i_class0 = np.where(target == 0)[0]
i_class1 = np.where(target == 1)[0]
# Number of observations in each class
n_class0 = len(i_class0)
n_class1 = len(i_class1)
# For every observation of class 0, randomly sample
# from class 1 without replacement
i_class1_downsampled = np.random.choice(i_class1, size=n_class0, replace=False)
# Join together class 0's target vector with the
# downsampled class 1's target vector
np.hstack((target[i_class0], target[i_class1_downsampled]))
array([0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1])
# Join together class 0's feature matrix with the
# downsampled class 1's feature matrix
np.vstack((features[i_class0,:], features[i_class1_downsampled,:]))[0:5]
array([[
[
[
[
[
5. ,
4.5,
4.4,
5. ,
5.1,
3.5,
2.3,
3.2,
3.5,
3.8,
1.3,
1.3,
1.3,
1.6,
1.9,
0.3],
0.3],
0.2],
0.6],
0.4]])
Our other option is to upsample the minority class. In upsampling, for every observa‐
tion in the majority class, we randomly select an observation from the minority class
with replacement. The end result is the same number of observations from the
92
|
Chapter 5: Handling Categorical Data
minority and majority classes. Upsampling is implemented very similarly to down‐
sampling, just in reverse:
# For every observation in class 1, randomly sample from class 0 with replacement
i_class0_upsampled = np.random.choice(i_class0, size=n_class1, replace=True)
# Join together class 0's upsampled target vector with class 1's target vector
np.concatenate((target[i_class0_upsampled], target[i_class1]))
array([0,
0,
0,
0,
0,
1,
1,
1,
1,
0,
0,
0,
0,
0,
1,
1,
1,
1,
0,
0,
0,
0,
0,
1,
1,
1,
1,
0,
0,
0,
0,
0,
1,
1,
1,
1,
0,
0,
0,
0,
0,
1,
1,
1,
1,
0,
0,
0,
0,
0,
1,
1,
1,
1,
0,
0,
0,
0,
0,
1,
1,
1,
1,
0,
0,
0,
0,
0,
1,
1,
1,
1,
0,
0,
0,
0,
1,
1,
1,
1,
1,
0,
0,
0,
0,
1,
1,
1,
1,
1,
0,
0,
0,
0,
1,
1,
1,
1,
1,
0,
0,
0,
0,
1,
1,
1,
1,
1,
0,
0,
0,
0,
1,
1,
1,
1,
1,
0,
0,
0,
0,
1,
1,
1,
1,
1,
0,
0,
0,
0,
1,
1,
1,
1,
1,
0, 0,
0, 0,
0, 0,
0, 0,
1, 1,
1, 1,
1, 1,
1, 1,
1])
0,
0,
0,
0,
1,
1,
1,
1,
0,
0,
0,
0,
1,
1,
1,
1,
0,
0,
0,
0,
1,
1,
1,
1,
0,
0,
0,
0,
1,
1,
1,
1,
0,
0,
0,
0,
1,
1,
1,
1,
0,
0,
0,
0,
1,
1,
1,
1,
# Join together class 0's upsampled feature matrix with class 1's feature matrix
np.vstack((features[i_class0_upsampled,:], features[i_class1,:]))[0:5]
array([[
[
[
[
[
5. ,
5. ,
5. ,
4.5,
4.8,
3.5,
3.5,
3.3,
2.3,
3. ,
1.6,
1.6,
1.4,
1.3,
1.4,
0.6],
0.6],
0.2],
0.3],
0.3]])
Discussion
In the real world, imbalanced classes are everywhere—most visitors don’t click the
buy button and many types of cancer are thankfully rare. For this reason, handling
imbalanced classes is a common activity in machine learning.
Our best strategy is simply to collect more observations—especially observations
from the minority class. However, this is often just not possible, so we have to resort
to other options.
A second strategy is to use a model evaluation metric better suited to imbalanced
classes. Accuracy is often used as a metric for evaluating the performance of a model,
but when imbalanced classes are present accuracy can be ill suited. For example, if
only 0.5% of observations have some rare cancer, then even a naive model that
predicts nobody has cancer will be 99.5% accurate. Clearly this is not ideal. Some bet‐
ter metrics we discuss in later chapters are confusion matrices, precision, recall, F1
scores, and ROC curves.
A third strategy is to use the class weighing parameters included in implementations
of some models. This allows us to have the algorithm adjust for imbalanced classes.
Fortunately, many scikit-learn classifiers have a class_weight parameter, making it a
good option.
5.5 Handling Imbalanced Classes
|
93
The fourth and fifth strategies are related: downsampling and upsampling. In down‐
sampling we create a random subset of the majority class of equal size to the minority
class. In upsampling we repeatedly sample with replacement from the minority class
to make it of equal size as the majority class. The decision between using downsam‐
pling and upsampling is context-specific, and in general we should try both to see
which produces better results.
94
|
Chapter 5: Handling Categorical Data
CHAPTER 6
Handling Text
6.0 Introduction
Unstructured text data, like the contents of a book or a tweet, is both one of the most
interesting sources of features and one of the most complex to handle. In this chapter,
we will cover strategies for transforming text into information-rich features. This is
not to say that the recipes covered here are comprehensive. There exist entire aca‐
demic disciplines focused on handling this and similar types of data, and the contents
of all their techniques would fill a small library. Despite this, there are some com‐
monly used techniques, and a knowledge of these will add valuable tools to our pre‐
processing toolbox.
6.1 Cleaning Text
Problem
You have some unstructured text data and want to complete some basic cleaning.
Solution
Most basic text cleaning operations should only replace Python’s core string opera‐
tions, in particular strip, replace, and split:
# Create text
text_data = ["
Interrobang. By Aishwarya Henriette
"Parking And Going. By Karl Gautier",
"
Today Is The night. By Jarek Prakash
",
"]
# Strip whitespaces
strip_whitespace = [string.strip() for string in text_data]
95
# Show text
strip_whitespace
['Interrobang. By Aishwarya Henriette',
'Parking And Going. By Karl Gautier',
'Today Is The night. By Jarek Prakash']
# Remove periods
remove_periods = [string.replace(".", "") for string in strip_whitespace]
# Show text
remove_periods
['Interrobang By Aishwarya Henriette',
'Parking And Going By Karl Gautier',
'Today Is The night By Jarek Prakash']
We also create and apply a custom transformation function:
# Create function
def capitalizer(string: str) -> str:
return string.upper()
# Apply function
[capitalizer(string) for string in remove_periods]
['INTERROBANG BY AISHWARYA HENRIETTE',
'PARKING AND GOING BY KARL GAUTIER',
'TODAY IS THE NIGHT BY JAREK PRAKASH']
Finally, we can use regular expressions to make powerful string operations:
# Import library
import re
# Create function
def replace_letters_with_X(string: str) -> str:
return re.sub(r"[a-zA-Z]", "X", string)
# Apply function
[replace_letters_with_X(string) for string in remove_periods]
['XXXXXXXXXXX XX XXXXXXXXX XXXXXXXXX',
'XXXXXXX XXX XXXXX XX XXXX XXXXXXX',
'XXXXX XX XXX XXXXX XX XXXXX XXXXXXX']
Discussion
Most text data will need to be cleaned before we can use it to build features. Most
basic text cleaning can be completed using Python’s standard string operations. In the
real world we will most likely define a custom cleaning function (e.g., capitalizer)
combining some cleaning tasks and apply that to the text data.
96
|
Chapter 6: Handling Text
See Also
• Beginners Tutorial for Regular Expressions in Python
6.2 Parsing and Cleaning HTML
Problem
You have text data with HTML elements and want to extract just the text.
Solution
Use Beautiful Soup’s extensive set of options to parse and extract from HTML:
# Load library
from bs4 import BeautifulSoup
# Create some HTML code
html = """
<div class='full_name'><span style='font-weight:bold'>
Masego</span> Azra</div>"
"""
# Parse html
soup = BeautifulSoup(html, "lxml")
# Find the div with the class "full_name", show text
soup.find("div", { "class" : "full_name" }).text
'Masego Azra'
Discussion
Despite the strange name, Beautiful Soup is a powerful Python library designed for
scraping HTML. Typically Beautiful Soup is used scrape live websites, but we can just
as easily use it to extract text data embedded in HTML. The full range of Beautiful
Soup operations is beyond the scope of this book, but even the few methods used in
our solution show how easily we can parse HTML code to extract the data we want.
See Also
• Beautiful Soup
6.2 Parsing and Cleaning HTML
|
97
6.3 Removing Punctuation
Problem
You have a feature of text data and want to remove punctuation.
Solution
Define a function that uses translate with a dictionary of punctuation characters:
# Load libraries
import unicodedata
import sys
# Create text
text_data = ['Hi!!!! I. Love. This. Song....',
'10000% Agree!!!! #LoveIT',
'Right?!?!']
# Create a dictionary of punctuation characters
punctuation = dict.fromkeys(i for i in range(sys.maxunicode)
if unicodedata.category(chr(i)).startswith('P'))
# For each string, remove any punctuation characters
[string.translate(punctuation) for string in text_data]
['Hi I Love This Song', '10000 Agree LoveIT', 'Right']
Discussion
translate is a Python method popular due to its blazing speed. In our solution, first
we created a dictionary, punctuation, with all punctuation characters according to
Unicode as its keys and None as its values. Next we translated all characters in the
string that are in punctuation into None, effectively removing them. There are more
readable ways to remove punctuation, but this somewhat hacky solution has the
advantage of being far faster than alternatives.
It is important to be conscious of the fact that punctuation contains information (e.g.,
“Right?” versus “Right!”). Removing punctuation is often a necessary evil to create
features; however, if the punctuation is important we should make sure to take that
into account.
6.4 Tokenizing Text
Problem
You have text and want to break it up into individual words.
98
|
Chapter 6: Handling Text
Solution
Natural Language Toolkit for Python (NLTK) has a powerful set of text manipulation
operations, including word tokenizing:
# Load library
from nltk.tokenize import word_tokenize
# Create text
string = "The science of today is the technology of tomorrow"
# Tokenize words
word_tokenize(string)
['The', 'science', 'of', 'today', 'is', 'the', 'technology', 'of', 'tomorrow']
We can also tokenize into sentences:
# Load library
from nltk.tokenize import sent_tokenize
# Create text
string = "The science of today is the technology of tomorrow. Tomorrow is today."
# Tokenize sentences
sent_tokenize(string)
['The science of today is the technology of tomorrow.', 'Tomorrow is today.']
Discussion
Tokenization, especially word tokenization, is a common task after cleaning text data
because it is the first step in the process of turning the text into data we will use to
construct useful features.
6.5 Removing Stop Words
Problem
Given tokenized text data, you want to remove extremely common words (e.g., a, is,
of, on) that contain little informational value.
Solution
Use NLTK’s stopwords:
# Load library
from nltk.corpus import stopwords
# You will have to download the set of stop words the first time
# import nltk
6.5 Removing Stop Words
|
99
# nltk.download('stopwords')
# Create word tokens
tokenized_words = ['i',
'am',
'going',
'to',
'go',
'to',
'the',
'store',
'and',
'park']
# Load stop words
stop_words = stopwords.words('english')
# Remove stop words
[word for word in tokenized_words if word not in stop_words]
['going', 'go', 'store', 'park']
Discussion
While “stop words” can refer to any set of words we want to remove before process‐
ing, frequently the term refers to extremely common words that themselves contain
little information value. NLTK has a list of common stop words that we can use to
find and remove stop words in our tokenized words:
# Show stop words
stop_words[:5]
['i', 'me', 'my', 'myself', 'we']
Note that NLTK’s stopwords assumes the tokenized words are all lowercased.
6.6 Stemming Words
Problem
You have tokenized words and want to convert them into their root forms.
Solution
Use NLTK’s PorterStemmer:
# Load library
from nltk.stem.porter import PorterStemmer
# Create word tokens
tokenized_words = ['i', 'am', 'humbled', 'by', 'this', 'traditional', 'meeting']
100
|
Chapter 6: Handling Text
# Create stemmer
porter = PorterStemmer()
# Apply stemmer
[porter.stem(word) for word in tokenized_words]
['i', 'am', 'humbl', 'by', 'thi', 'tradit', 'meet']
Discussion
Stemming reduces a word to its stem by identifying and removing affixes (e.g., ger‐
unds) while keeping the root meaning of the word. For example, both “tradition” and
“traditional” have “tradit” as their stem, indicating that while they are different words
they represent the same general concept. By stemming our text data, we transform it
to something less readable, but closer to its base meaning and thus more suitable for
comparison across observations. NLTK’s PorterStemmer implements the widely used
Porter stemming algorithm to remove or replace common suffixes to produce the
word stem.
See Also
• Porter Stemming Algorithm
6.7 Tagging Parts of Speech
Problem
You have text data and want to tag each word or character with its part of speech.
Solution
Use NLTK’s pre-trained parts-of-speech tagger:
# Load libraries
from nltk import pos_tag
from nltk import word_tokenize
# Create text
text_data = "Chris loved outdoor running"
# Use pre-trained part of speech tagger
text_tagged = pos_tag(word_tokenize(text_data))
# Show parts of speech
text_tagged
[('Chris', 'NNP'), ('loved', 'VBD'), ('outdoor', 'RP'), ('running', 'VBG')]
6.7 Tagging Parts of Speech
|
101
The output is a list of tuples with the word and the tag of the part of speech. NLTK
uses the Penn Treebank parts for speech tags. Some examples of the Penn Treebank
tags are:
Tag
NNP
NN
RB
VBD
VBG
JJ
PRP
Part of speech
Proper noun, singular
Noun, singular or mass
Adverb
Verb, past tense
Verb, gerund or present participle
Adjective
Personal pronoun
Once the text has been tagged, we can use the tags to find certain parts of speech. For
example, here are all nouns:
# Filter words
[word for word, tag in text_tagged if tag in ['NN','NNS','NNP','NNPS'] ]
['Chris']
A more realistic situation would be that we have data where every observation con‐
tains a tweet and we want to convert those sentences into features for individual parts
of speech (e.g., a feature with 1 if a proper noun is present, and 0 otherwise):
# Create text
tweets = ["I am eating a burrito for breakfast",
"Political science is an amazing field",
"San Francisco is an awesome city"]
# Create list
tagged_tweets = []
# Tag each word and each tweet
for tweet in tweets:
tweet_tag = nltk.pos_tag(word_tokenize(tweet))
tagged_tweets.append([tag for word, tag in tweet_tag])
# Use one-hot encoding to convert the tags into features
one_hot_multi = MultiLabelBinarizer()
one_hot_multi.fit_transform(tagged_tweets)
array([[1, 1, 0, 1, 0, 1, 1, 1, 0],
[1, 0, 1, 1, 0, 0, 0, 0, 1],
[1, 0, 1, 1, 1, 0, 0, 0, 1]])
Using classes_ we can see that each feature is a part-of-speech tag:
# Show feature names
one_hot_multi.classes_
102
|
Chapter 6: Handling Text
array(['DT', 'IN', 'JJ', 'NN', 'NNP', 'PRP', 'VBG', 'VBP', 'VBZ'], dtype=object)
Discussion
If our text is English and not on a specialized topic (e.g., medicine) the simplest solu‐
tion is to use NLTK’s pre-trained parts-of-speech tagger. However, if pos_tag is not
very accurate, NLTK also gives us the ability to train our own tagger. The major
downside of training a tagger is that we need a large corpus of text where the tag of
each word is known. Constructing this tagged corpus is obviously labor intensive and
is probably going to be a last resort.
All that said, if we had a tagged corpus and wanted to train a tagger, the following is
an example of how we could do it. The corpus we are using is the Brown Corpus, one
of the most popular sources of tagged text. Here we use a backoff n-gram tagger,
where n is the number of previous words we take into account when predicting a
word’s part-of-speech tag. First we take into account the previous two words using
TrigramTagger; if two words are not present, we “back off ” and take into account the
tag of the previous one word using BigramTagger, and finally if that fails we only look
at the word itself using UnigramTagger. To examine the accuracy of our tagger, we
split our text data into two parts, train our tagger on one part, and test how well it
predicts the tags of the second part:
# Load library
from nltk.corpus import brown
from nltk.tag import UnigramTagger
from nltk.tag import BigramTagger
from nltk.tag import TrigramTagger
# Get some text from the Brown Corpus, broken into sentences
sentences = brown.tagged_sents(categories='news')
# Split into 4000 sentences for training and 623 for testing
train = sentences[:4000]
test = sentences[4000:]
# Create backoff tagger
unigram = UnigramTagger(train)
bigram = BigramTagger(train, backoff=unigram)
trigram = TrigramTagger(train, backoff=bigram)
# Show accuracy
trigram.evaluate(test)
0.8179229731754832
See Also
• Penn Treebank
6.7 Tagging Parts of Speech
|
103
• Brown Corpus
6.8 Encoding Text as a Bag of Words
Problem
You have text data and want to create a set of features indicating the number of times
an observation’s text contains a particular word.
Solution
Use scikit-learn’s CountVectorizer:
# Load library
import numpy as np
from sklearn.feature_extraction.text import CountVectorizer
# Create text
text_data = np.array(['I love Brazil. Brazil!',
'Sweden is best',
'Germany beats both'])
# Create the bag of words feature matrix
count = CountVectorizer()
bag_of_words = count.fit_transform(text_data)
# Show feature matrix
bag_of_words
<3x8 sparse matrix of type '<class 'numpy.int64'>'
with 8 stored elements in Compressed Sparse Row format>
This output is a sparse array, which is often necessary when we have a large amount
of text. However, in our toy example we can use toarray to view a matrix of word
counts for each observation:
bag_of_words.toarray()
array([[0, 0, 0, 2, 0, 0, 1, 0],
[0, 1, 0, 0, 0, 1, 0, 1],
[1, 0, 1, 0, 1, 0, 0, 0]], dtype=int64)
We can use the vocabulary_ method to view the word associated with each feature:
# Show feature names
count.get_feature_names()
['beats', 'best', 'both', 'brazil', 'germany', 'is', 'love', 'sweden']
104
|
Chapter 6: Handling Text
This might be confusing, so for the sake of clarity here is what the feature matrix
looks like with the words as column names (each row is one observation):
beats
0
0
1
best
0
1
0
both
0
0
1
brazil
2
0
0
germany
0
0
1
is
0
1
0
love
1
0
0
sweden
0
1
0
Discussion
One of the most common methods of transforming text into features is by using a
bag-of-words model. Bag-of-words models output a feature for every unique word in
text data, with each feature containing a count of occurrences in observations. For
example, in our solution the sentence I love Brazil. Brazil! has a value of 2 in
the “brazil” feature because the word brazil appears two times.
The text data in our solution was purposely small. In the real world, a single observa‐
tion of text data could be the contents of an entire book! Since our bag-of-words
model creates a feature for every unique word in the data, the resulting matrix can
contain thousands of features. This means that the size of the matrix can sometimes
become very large in memory. However, luckily we can exploit a common character‐
istic of bag-of-words feature matrices to reduce the amount of data we need to store.
Most words likely do not occur in most observations, and therefore bag-of-words fea‐
ture matrices will contain mostly 0s as values. We call these types of matrices “sparse.”
Instead of storing all values of the matrix, we can only store nonzero values and then
assume all other values are 0. This will save us memory when we have large feature
matrices. One of the nice features of CountVectorizer is that the output is a sparse
matrix by default.
CountVectorizer comes with a number of useful parameters to make creating bagof-words feature matrices easy. First, while by default every feature is a word, that
does not have to be the case. Instead we can set every feature to be the combination of
two words (called a 2-gram) or even three words (3-gram). ngram_range sets the
minimum and maximum size of our n-grams. For example, (2,3) will return all 2grams and 3-grams. Second, we can easily remove low-information filler words using
stop_words either with a built-in list or a custom list. Finally, we can restrict the
words or phrases we want to consider to a certain list of words using vocabulary. For
example, we could create a bag-of-words feature matrix for only occurrences of coun‐
try names:
# Create feature matrix with arguments
count_2gram = CountVectorizer(ngram_range=(1,2),
stop_words="english",
vocabulary=['brazil'])
6.8 Encoding Text as a Bag of Words
|
105
bag = count_2gram.fit_transform(text_data)
# View feature matrix
bag.toarray()
array([[2],
[0],
[0]])
# View the 1-grams and 2-grams
count_2gram.vocabulary_
{'brazil': 0}
See Also
• n-gram
• Bag of Words Meets Bags of Popcorn
6.9 Weighting Word Importance
Problem
You want a bag of words, but with words weighted by their importance to an observa‐
tion.
Solution
Compare the frequency of the word in a document (a tweet, movie review, speech
transcript, etc.) with the frequency of the word in all other documents using term
frequency-inverse document frequency (tf-idf). scikit-learn makes this easy with
TfidfVectorizer:
# Load libraries
import numpy as np
from sklearn.feature_extraction.text import TfidfVectorizer
# Create text
text_data = np.array(['I love Brazil. Brazil!',
'Sweden is best',
'Germany beats both'])
# Create the tf-idf feature matrix
tfidf = TfidfVectorizer()
feature_matrix = tfidf.fit_transform(text_data)
# Show tf-idf feature matrix
feature_matrix
106
|
Chapter 6: Handling Text
<3x8 sparse matrix of type '<class 'numpy.float64'>'
with 8 stored elements in Compressed Sparse Row format>
Just as in Recipe 6.8, the output is a spare matrix. However, if we want to view the
output as a dense matrix, we can use .toarray:
# Show tf-idf feature matrix as dense matrix
feature_matrix.toarray()
array([[ 0.
,
0.
,
[ 0.
,
0.57735027,
[ 0.57735027,
0.
,
0.
,
0.4472136 ,
0.57735027,
0.
,
0.
,
0.
,
0.
, 0.89442719,
0.
],
0.
, 0.
,
0.57735027],
0.57735027, 0.
,
0.
]])
0.
,
0.
,
0.57735027,
vocabulary_ shows us the word of each feature:
# Show feature names
tfidf.vocabulary_
{'beats': 0,
'best': 1,
'both': 2,
'brazil': 3,
'germany': 4,
'is': 5,
'love': 6,
'sweden': 7}
Discussion
The more a word appears in a document, the more likely it is important to that docu‐
ment. For example, if the word economy appears frequently, it is evidence that the
document might be about economics. We call this term frequency (tf).
In contrast, if a word appears in many documents, it is likely less important to any
individual document. For example, if every document in some text data contains the
word after then it is probably an unimportant word. We call this document frequency
(df).
By combining these two statistics, we can assign a score to every word representing
how important that word is in a document. Specifically, we multiply tf to the inverse
of document frequency (idf):
tf‐idf t, d = t f t, d × id f t
where t is a word and d is a document. There are a number of variations in how tf and
idf are calculated. In scikit-learn, tf is simply the number of times a word appears in
the document and idf is calculated as:
6.9 Weighting Word Importance
|
107
idf t = log
1 + nd
+1
1 + df d, t
where nd is the number of documents and df(d,t) is term, t’s document frequency (i.e.,
number of documents where the term appears).
By default, scikit-learn then normalizes the tf-idf vectors using the Euclidean norm
(L2 norm). The higher the resulting value, the more important the word is to a docu‐
ment.
See Also
• scikit-learn documentation: tf–idf term weighting
108
|
Chapter 6: Handling Text
CHAPTER 7
Handling Dates and Times
7.0 Introduction
Dates and times (datetimes) are frequently encountered during preprocessing for
machine learning, whether the time of a particular sale or the year of some public
health statistic. In this chapter, we will build a toolbox of strategies for handling time
series data including tackling time zones and creating lagged time features. Specifi‐
cally, we will focus on the time series tools in the pandas library, which centralizes the
functionality of many other libraries.
7.1 Converting Strings to Dates
Problem
Given a vector of strings representing dates and times, you want to transform them
into time series data.
Solution
Use pandas’ to_datetime with the format of the date and/or time specified in the
format parameter:
# Load libraries
import numpy as np
import pandas as pd
# Create strings
date_strings = np.array(['03-04-2005 11:35 PM',
'23-05-2010 12:01 AM',
'04-09-2009 09:09 PM'])
109
# Convert to datetimes
[pd.to_datetime(date, format='%d-%m-%Y %I:%M %p') for date in date_strings]
[Timestamp('2005-04-03 23:35:00'),
Timestamp('2010-05-23 00:01:00'),
Timestamp('2009-09-04 21:09:00')]
We might also want to add an argument to the errors parameter to handle problems:
# Convert to datetimes
[pd.to_datetime(date, format="%d-%m-%Y %I:%M %p", errors="coerce")
for date in date_strings]
[Timestamp('2005-04-03 23:35:00'),
Timestamp('2010-05-23 00:01:00'),
Timestamp('2009-09-04 21:09:00')]
If errors="coerce", then any problem that occurs will not raise an error (the default
behavior) but instead will set the value causing the error to NaT (i.e., a missing value).
Discussion
When dates and times come as strings, we need to convert them into a data type
Python can understand. While there are a number of Python tools for converting
strings to datetimes, following our use of pandas in other recipes we can use to_date
time to conduct the transformation. One obstacle to strings representing dates and
times is that the format of the strings can vary significantly between data sources. For
example, one vector of dates might represent March 23rd, 2015 as “03-23-15” while
another might use “3|23|2015”. We can use the format parameter to specify the exact
format of the string. Here are some common date and time formatting codes:
Code Description
Full year
%Y
Month w/ zero padding
%m
Day of the month w/ zero padding
%d
Hour (12hr clock) w/ zero padding
%I
AM or PM
%p
Minute w/ zero padding
%M
Second w/ zero padding
%S
Example
2001
04
09
02
AM
05
09
See Also
• Complete List of Python String Time Codes
110
|
Chapter 7: Handling Dates and Times
7.2 Handling Time Zones
Problem
You have time series data and want to add or change time zone information.
Solution
If not specified, pandas objects have no time zone. However, we can add a time zone
using tz during creation:
# Load library
import pandas as pd
# Create datetime
pd.Timestamp('2017-05-01 06:00:00', tz='Europe/London')
Timestamp('2017-05-01 06:00:00+0100', tz='Europe/London')
We can add a time zone to a previously created datetime using tz_localize:
# Create datetime
date = pd.Timestamp('2017-05-01 06:00:00')
# Set time zone
date_in_london = date.tz_localize('Europe/London')
# Show datetime
date_in_london
Timestamp('2017-05-01 06:00:00+0100', tz='Europe/London')
We can also convert to a different time zone:
# Change time zone
date_in_london.tz_convert('Africa/Abidjan')
Timestamp('2017-05-01 05:00:00+0000', tz='Africa/Abidjan')
Finally, pandas’ Series objects can apply tz_localize and tz_convert to every ele‐
ment:
# Create three dates
dates = pd.Series(pd.date_range('2/2/2002', periods=3, freq='M'))
# Set time zone
dates.dt.tz_localize('Africa/Abidjan')
0
2002-02-28 00:00:00+00:00
1
2002-03-31 00:00:00+00:00
2
2002-04-30 00:00:00+00:00
dtype: datetime64[ns, Africa/Abidjan]
7.2 Handling Time Zones
|
111
Discussion
pandas supports two sets of strings representing timezones; however, I suggest using
pytz library’s strings. We can see all the strings used to represent time zones by
importing all_timezones:
# Load library
from pytz import all_timezones
# Show two time zones
all_timezones[0:2]
['Africa/Abidjan', 'Africa/Accra']
7.3 Selecting Dates and Times
Problem
You have a vector of dates and you want to select one or more.
Solution
Use two boolean conditions as the start and end dates:
# Load library
import pandas as pd
# Create data frame
dataframe = pd.DataFrame()
# Create datetimes
dataframe['date'] = pd.date_range('1/1/2001', periods=100000, freq='H')
# Select observations between two datetimes
dataframe[(dataframe['date'] > '2002-1-1 01:00:00') &
(dataframe['date'] <= '2002-1-1 04:00:00')]
date
8762 2002-01-01 02:00:00
8763 2002-01-01 03:00:00
8764 2002-01-01 04:00:00
Alternatively, we can set the date column as the DataFrame’s index and then slice
using loc:
# Set index
dataframe = dataframe.set_index(dataframe['date'])
112
|
Chapter 7: Handling Dates and Times
# Select observations between two datetimes
dataframe.loc['2002-1-1 01:00:00':'2002-1-1 04:00:00']
2002-01-01 01:00:00
2002-01-01 02:00:00
2002-01-01 03:00:00
2002-01-01 04:00:00
date
date
2002-01-01 01:00:00
2002-01-01 02:00:00
2002-01-01 03:00:00
2002-01-01 04:00:00
Discussion
Whether we use boolean conditions or index slicing is situation dependent. If we
wanted to do some complex time series manipulation, it might be worth the overhead
of setting the date column as the index of the DataFrame, but if we wanted to do
some simple data wrangling, the boolean conditions might be easier.
7.4 Breaking Up Date Data into Multiple Features
Problem
You have a column of dates and times and you want to create features for year,
month, day, hour, and minute.
Solution
Use pandas Series.dt’s time properties:
# Load library
import pandas as pd
# Create data frame
dataframe = pd.DataFrame()
# Create five dates
dataframe['date'] = pd.date_range('1/1/2001', periods=150, freq='W')
# Create features for year, month, day, hour, and minute
dataframe['year'] = dataframe['date'].dt.year
dataframe['month'] = dataframe['date'].dt.month
dataframe['day'] = dataframe['date'].dt.day
dataframe['hour'] = dataframe['date'].dt.hour
dataframe['minute'] = dataframe['date'].dt.minute
# Show three rows
dataframe.head(3)
7.4 Breaking Up Date Data into Multiple Features
|
113
date
year month
0 2001-01-07 2001 1
1 2001-01-14 2001 1
2 2001-01-21 2001 1
day
7
14
21
hour
0
0
0
minute
0
0
0
Discussion
Sometimes it can be useful to break up a column of dates into components. For
example, we might want a feature that just includes the year of the observation or we
might want only to consider the month of some observation so we can compare them
regardless of year.
7.5 Calculating the Difference Between Dates
Problem
You have two datetime features and want to calculate the time between them for each
observation.
Solution
Subtract the two date features using pandas:
# Load library
import pandas as pd
# Create data frame
dataframe = pd.DataFrame()
# Create two datetime features
dataframe['Arrived'] = [pd.Timestamp('01-01-2017'), pd.Timestamp('01-04-2017')]
dataframe['Left'] = [pd.Timestamp('01-01-2017'), pd.Timestamp('01-06-2017')]
# Calculate duration between features
dataframe['Left'] - dataframe['Arrived']
0
0 days
1
2 days
dtype: timedelta64[ns]
Often we will want to remove the days output and keep only the numerical value:
# Calculate duration between features
pd.Series(delta.days for delta in (dataframe['Left'] - dataframe['Arrived']))
0
0
1
2
dtype: int64
114
|
Chapter 7: Handling Dates and Times
Discussion
There are times when the feature we want is the change (delta) between two points in
time. For example, we might have the dates a customer checks in and checks out of a
hotel, but the feature we want is the duration of his stay. pandas makes this calcula‐
tion easy using the TimeDelta data type.
See Also
• Pandas documentation: Time Deltas
7.6 Encoding Days of the Week
Problem
You have a vector of dates and want to know the day of the week for each date.
Solution
Use pandas’ Series.dt property weekday_name:
# Load library
import pandas as pd
# Create dates
dates = pd.Series(pd.date_range("2/2/2002", periods=3, freq="M"))
# Show days of the week
dates.dt.weekday_name
0
Thursday
1
Sunday
2
Tuesday
dtype: object
If we want the output to be a numerical value and therefore more usable as a machine
learning feature, we can use weekday where the days of the week are represented as an
integer (Monday is 0):
# Show days of the week
dates.dt.weekday
0
3
1
6
2
1
dtype: int64
7.6 Encoding Days of the Week
|
115
Discussion
Knowing the weekday can be helpful if, for instance, we wanted to compare total sales
on Sundays for the past three years. pandas makes creating a feature vector contain‐
ing weekday information easy.
See Also
• pandas Series datetimelike properties
7.7 Creating a Lagged Feature
Problem
You want to create a feature that is lagged n time periods.
Solution
Use pandas’ shift:
# Load library
import pandas as pd
# Create data frame
dataframe = pd.DataFrame()
# Create data
dataframe["dates"] = pd.date_range("1/1/2001", periods=5, freq="D")
dataframe["stock_price"] = [1.1,2.2,3.3,4.4,5.5]
# Lagged values by one row
dataframe["previous_days_stock_price"] = dataframe["stock_price"].shift(1)
# Show data frame
dataframe
0
1
2
3
4
116
dates
2001-01-01
2001-01-02
2001-01-03
2001-01-04
2001-01-05
|
stock_price
1.1
2.2
3.3
4.4
5.5
previous_days_stock_price
NaN
1.1
2.2
3.3
4.4
Chapter 7: Handling Dates and Times
Discussion
Very often data is based on regularly spaced time periods (e.g., every day, every hour,
every three hours) and we are interested in using values in the past to make predic‐
tions (this is often called lagging a feature). For example, we might want to predict a
stock’s price using the price it was the day before. With pandas we can use shift to
lag values by one row, creating a new feature containing past values.
In our solution, the first row for previous_days_stock_price is a missing value
because there is no previous stock_price value.
7.8 Using Rolling Time Windows
Problem
Given time series data, you want to calculate some statistic for a rolling time.
Solution
# Load library
import pandas as pd
# Create datetimes
time_index = pd.date_range("01/01/2010", periods=5, freq="M")
# Create data frame, set index
dataframe = pd.DataFrame(index=time_index)
# Create feature
dataframe["Stock_Price"] = [1,2,3,4,5]
# Calculate rolling mean
dataframe.rolling(window=2).mean()
2010-01-31
2010-02-28
2010-03-31
2010-04-30
2010-05-31
Stock_Price
NaN
1.5
2.5
3.5
4.5
Discussion
Rolling (also called moving) time windows are conceptually simple but can be diffi‐
cult to understand at first. Imagine we have monthly observations for a stock’s price.
It is often useful to have a time window of a certain number of months and then
7.8 Using Rolling Time Windows
|
117
move over the observations calculating a statistic for all observations in the time win‐
dow.
For example, if we have a time window of three months and we want a rolling mean,
we would calculate:
1. mean(January, February, March)
2. mean(February, March, April)
3. mean(March, April, May)
4. etc.
Another way to put it: our three-month time window “walks” over the observations,
calculating the window’s mean at each step.
pandas’ rolling allows us to specify the size of the window using window and then
quickly calculate some common statistics, including the max value (max()), mean
value (mean()), count of values (count()), and rolling correlation (corr()).
Rolling means are often used to smooth out time series data because using the mean
of the entire time window dampens the effect of short-term fluctuations.
See Also
• pandas documentation: Rolling Windows
• What are Moving Average or Smoothing Techniques?
7.9 Handling Missing Data in Time Series
Problem
You have missing values in time series data.
Solution
In addition to the missing data strategies previously discussed, when we have time
series data we can use interpolation to fill in gaps caused by missing values:
# Load libraries
import pandas as pd
import numpy as np
# Create date
time_index = pd.date_range("01/01/2010", periods=5, freq="M")
# Create data frame, set index
118
|
Chapter 7: Handling Dates and Times
dataframe = pd.DataFrame(index=time_index)
# Create feature with a gap of missing values
dataframe["Sales"] = [1.0,2.0,np.nan,np.nan,5.0]
# Interpolate missing values
dataframe.interpolate()
2010-01-31
2010-02-28
2010-03-31
2010-04-30
2010-05-31
Sales
1.0
2.0
3.0
4.0
5.0
Alternatively, we can replace missing values with the last known value (i.e., forwardfilling):
# Forward-fill
dataframe.ffill()
2010-01-31
2010-02-28
2010-03-31
2010-04-30
2010-05-31
Sales
1.0
2.0
2.0
2.0
5.0
We can also replace missing values with the latest known value (i.e., back-filling):
# Back-fill
dataframe.bfill()
2010-01-31
2010-02-28
2010-03-31
2010-04-30
2010-05-31
Sales
1.0
2.0
5.0
5.0
5.0
Discussion
Interpolation is a technique for filling in gaps caused by missing values by, in effect,
drawing a line or curve between the known values bordering the gap and using that
line or curve to predict reasonable values. Interpolation can be particularly useful
7.9 Handling Missing Data in Time Series
|
119
when the time intervals between are constant, the data is not prone to noisy fluctua‐
tions, and the gaps caused by missing values are small. For example, in our solution a
gap of two missing values was bordered by 2.0 and 5.0. By fitting a line starting at
2.0 and ending at 5.0, we can make reasonable guesses for the two missing values in
between of 3.0 and 4.0.
If we believe the line between the two known points is nonlinear, we can use interpo
late’s method to specify the interpolation method:
# Interpolate missing values
dataframe.interpolate(method="quadratic")
2010-01-31
2010-02-28
2010-03-31
2010-04-30
2010-05-31
Sales
1.000000
2.000000
3.059808
4.038069
5.000000
Finally, there might be cases when we have large gaps of missing values and do not
want to interpolate values across the entire gap. In these cases we can use limit to
restrict the number of interpolated values and limit_direction to set whether to
interpolate values forward from at the last known value before the gap or vice versa:
# Interpolate missing values
dataframe.interpolate(limit=1, limit_direction="forward")
2010-01-31
2010-02-28
2010-03-31
2010-04-30
2010-05-31
Sales
1.0
2.0
3.0
NaN
5.0
Back-filling and forward-filling can be thought of as a form of naive interpolation,
where we draw a flat line from a known value and use it to fill in missing values. One
(minor) advantage back- and forward-filling have over interpolation is the lack of the
need for known values on both sides of missing value(s).
120
|
Chapter 7: Handling Dates and Times
CHAPTER 8
Handling Images
8.0 Introduction
Image classification is one of the most exciting areas of machine learning. The ability
of computers to recognize patterns and objects from images is an incredibly powerful
tool in our toolkit. However, before we can apply machine learning to images, we
often first need to transform the raw images to features usable by our learning algo‐
rithms.
To work with images, we will use the Open Source Computer Vision Library
(OpenCV). While there are a number of good libraries out there, OpenCV is the
most popular and documented library for handling images. One of the biggest hur‐
dles to using OpenCV is installing it. However, fortunately if we are using Python 3
(at the time of publication OpenCV does not work with Python 3.6+), we can use
Anaconda’s package manager tool conda to install OpenCV in a single line of code in
our terminal:
conda install --channel https://conda.anaconda.org/menpo opencv3
Afterward, we can check the installation by opening a notebook, importing OpenCV,
and checking the version number (3.1.0):
import cv2
cv2.__version__
If installing OpenCV using conda does not work, there are many guides online.
Finally, throughout this chapter we will use a set of images as examples, which are
available to download on GitHub.
121
8.1 Loading Images
Problem
You want to load an image for preprocessing.
Solution
Use OpenCV’s imread:
# Load library
import cv2
import numpy as np
from matplotlib import pyplot as plt
# Load image as grayscale
image = cv2.imread("images/plane.jpg", cv2.IMREAD_GRAYSCALE)
If we want to view the image, we can use the Python plotting library Matplotlib:
# Show image
plt.imshow(image, cmap="gray"), plt.axis("off")
plt.show()
122
|
Chapter 8: Handling Images
Discussion
Fundamentally, images are data and when we use imread we convert that data into a
data type we are very familiar with—a NumPy array:
# Show data type
type(image)
numpy.ndarray
We have transformed the image into a matrix whose elements correspond to individ‐
ual pixels. We can even take a look at the actual values of the matrix:
# Show image data
image
array([[140,
[144,
[152,
...,
[156,
[146,
[143,
136, 146, ..., 132, 139, 134],
136, 149, ..., 142, 124, 126],
139, 144, ..., 121, 127, 134],
146, 144, ..., 157, 154, 151],
150, 147, ..., 156, 158, 157],
138, 147, ..., 156, 157, 157]], dtype=uint8)
The resolution of our image was 3600 × 2270, the exact dimensions of our matrix:
# Show dimensions
image.shape
(2270, 3600)
What does each element in the matrix actually represent? In grayscale images, the
value of an individual element is the pixel intensity. Intensity values range from black
(0) to white (255). For example, the intensity of the top-rightmost pixel in our image
has a value of 140:
# Show first pixel
image[0,0]
140
In the matrix, each element contains three values corresponding to blue, green, red
values (BGR):
# Load image in color
image_bgr = cv2.imread("images/plane.jpg", cv2.IMREAD_COLOR)
# Show pixel
image_bgr[0,0]
array([195, 144, 111], dtype=uint8)
One small caveat: by default OpenCV uses BGR, but many image applications—
including Matplotlib—use red, green, blue (RGB), meaning the red and the blue
8.1 Loading Images
|
123
values are swapped. To properly display OpenCV color images in Matplotlib, we need
to first convert the color to RGB (apologies to hardcopy readers):
# Convert to RGB
image_rgb = cv2.cvtColor(image_bgr, cv2.COLOR_BGR2RGB)
# Show image
plt.imshow(image_rgb), plt.axis("off")
plt.show()
See Also
• Difference between RGB and BGR
• RGB color model
8.2 Saving Images
Problem
You want to save an image for preprocessing.
Solution
Use OpenCV’s imwrite:
124
|
Chapter 8: Handling Images
# Load libraries
import cv2
import numpy as np
from matplotlib import pyplot as plt
# Load image as grayscale
image = cv2.imread("images/plane.jpg", cv2.IMREAD_GRAYSCALE)
# Save image
cv2.imwrite("images/plane_new.jpg", image)
True
Discussion
OpenCV’s imwrite saves images to the filepath specified. The format of the image is
defined by the filename’s extension (.jpg, .png, etc.). One behavior to be careful about:
imwrite will overwrite existing files without outputting an error or asking for confir‐
mation.
8.3 Resizing Images
Problem
You want to resize an image for further preprocessing.
Solution
Use resize to change the size of an image:
# Load image
import cv2
import numpy as np
from matplotlib import pyplot as plt
# Load image as grayscale
image = cv2.imread("images/plane_256x256.jpg", cv2.IMREAD_GRAYSCALE)
# Resize image to 50 pixels by 50 pixels
image_50x50 = cv2.resize(image, (50, 50))
# View image
plt.imshow(image_50x50, cmap="gray"), plt.axis("off")
plt.show()
8.3 Resizing Images
|
125
Discussion
Resizing images is a common task in image preprocessing for two reasons. First,
images come in all shapes and sizes, and to be usable as features, images must have
the same dimensions. This standardization of image size does come with costs, how‐
ever; images are matrices of information and when we reduce the size of the image
we are reducing the size of that matrix and the information it contains. Second,
machine learning can require thousands or hundreds of thousands of images. When
those images are very large they can take up a lot of memory, and by resizing them we
can dramatically reduce memory usage. Some common image sizes for machine
learning are 32 × 32, 64 × 64, 96 × 96, and 256 × 256.
8.4 Cropping Images
Problem
You want to remove the outer portion of the image to change its dimensions.
126
|
Chapter 8: Handling Images
Solution
The image is encoded as a two-dimensional NumPy array, so we can crop the image
easily by slicing the array:
# Load image
import cv2
import numpy as np
from matplotlib import pyplot as plt
# Load image in grayscale
image = cv2.imread("images/plane_256x256.jpg", cv2.IMREAD_GRAYSCALE)
# Select first half of the columns and all rows
image_cropped = image[:,:128]
# Show image
plt.imshow(image_cropped, cmap="gray"), plt.axis("off")
plt.show()
Discussion
Since OpenCV represents images as a matrix of elements, by selecting the rows and
columns we want to keep we are able to easily crop the image. Cropping can be par‐
ticularly useful if we know that we only want to keep a certain part of every image.
8.4 Cropping Images
|
127
For example, if our images come from a stationary security camera we can crop all
the images so they only contain the area of interest.
See Also
• Slicing NumPy Arrays
8.5 Blurring Images
Problem
You want to smooth out an image.
Solution
To blur an image, each pixel is transformed to be the average value of its neighbors.
This neighbor and the operation performed are mathematically represented as a ker‐
nel (don’t worry if you don’t know what a kernel is). The size of this kernel deter‐
mines the amount of blurring, with larger kernels producing smoother images. Here
we blur an image by averaging the values of a 5 × 5 kernel around each pixel:
# Load libraries
import cv2
import numpy as np
from matplotlib import pyplot as plt
# Load image as grayscale
image = cv2.imread("images/plane_256x256.jpg", cv2.IMREAD_GRAYSCALE)
# Blur image
image_blurry = cv2.blur(image, (5,5))
# Show image
plt.imshow(image_blurry, cmap="gray"), plt.axis("off")
plt.show()
128
|
Chapter 8: Handling Images
To highlight the effect of kernel size, here is the same blurring with a 100 × 100
kernel:
# Blur image
image_very_blurry = cv2.blur(image, (100,100))
# Show image
plt.imshow(image_very_blurry, cmap="gray"), plt.xticks([]), plt.yticks([])
plt.show()
8.5 Blurring Images
|
129
Discussion
Kernels are widely used in image processing to do everything from sharpening to
edge detection, and will come up repeatedly in this chapter. The blurring kernel we
used looks like this:
# Create kernel
kernel = np.ones((5,5)) / 25.0
# Show kernel
kernel
array([[
[
[
[
[
0.04,
0.04,
0.04,
0.04,
0.04,
0.04,
0.04,
0.04,
0.04,
0.04,
0.04,
0.04,
0.04,
0.04,
0.04,
0.04,
0.04,
0.04,
0.04,
0.04,
0.04],
0.04],
0.04],
0.04],
0.04]])
The center element in the kernel is the pixel being examined, while the remaining ele‐
ments are its neighbors. Since all elements have the same value (normalized to add up
to 1), each has an equal say in the resulting value of the pixel of interest. We can man‐
ually apply a kernel to an image using filter2D to produce a similar blurring effect:
# Apply kernel
image_kernel = cv2.filter2D(image, -1, kernel)
# Show image
130
|
Chapter 8: Handling Images
plt.imshow(image_kernel, cmap="gray"), plt.xticks([]), plt.yticks([])
plt.show()
See Also
• Image Kernels Explained Visually
• Common Image Kernels
8.6 Sharpening Images
Problem
You want to sharpen an image.
Solution
Create a kernel that highlights the target pixel. Then apply it to the image using fil
ter2D:
# Load libraries
import cv2
import numpy as np
from matplotlib import pyplot as plt
8.6 Sharpening Images
|
131
# Load image as grayscale
image = cv2.imread("images/plane_256x256.jpg", cv2.IMREAD_GRAYSCALE)
# Create kernel
kernel = np.array([[0, -1, 0],
[-1, 5,-1],
[0, -1, 0]])
# Sharpen image
image_sharp = cv2.filter2D(image, -1, kernel)
# Show image
plt.imshow(image_sharp, cmap="gray"), plt.axis("off")
plt.show()
Discussion
Sharpening works similarly to blurring, except instead of using a kernel to average
the neighboring values, we constructed a kernel to highlight the pixel itself. The
resulting effect makes contrasts in edges stand out more in the image.
132
|
Chapter 8: Handling Images
8.7 Enhancing Contrast
Problem
We want to increase the contrast between pixels in an image.
Solution
Histogram equalization is a tool for image processing that can make objects and
shapes stand out. When we have a grayscale image, we can apply OpenCV’s equali
zeHist directly on the image:
# Load libraries
import cv2
import numpy as np
from matplotlib import pyplot as plt
# Load image
image = cv2.imread("images/plane_256x256.jpg", cv2.IMREAD_GRAYSCALE)
# Enhance image
image_enhanced = cv2.equalizeHist(image)
# Show image
plt.imshow(image_enhanced, cmap="gray"), plt.axis("off")
plt.show()
8.7 Enhancing Contrast
|
133
However, when we have a color image, we first need to convert the image to the YUV
color format. The Y is the luma, or brightness, and U and V denote the color. After
the conversion, we can apply equalizeHist to the image and then convert it back to
BGR or RGB:
# Load image
image_bgr = cv2.imread("images/plane.jpg")
# Convert to YUV
image_yuv = cv2.cvtColor(image_bgr, cv2.COLOR_BGR2YUV)
# Apply histogram equalization
image_yuv[:, :, 0] = cv2.equalizeHist(image_yuv[:, :, 0])
# Convert to RGB
image_rgb = cv2.cvtColor(image_yuv, cv2.COLOR_YUV2RGB)
# Show image
plt.imshow(image_rgb), plt.axis("off")
plt.show()
134
|
Chapter 8: Handling Images
Discussion
While a detailed explanation of how histogram equalization works is beyond the
scope of this book, the short explanation is that it transforms the image so that it uses
a wider range of pixel intensities.
While the resulting image often does not look “realistic,” we need to remember that
the image is just a visual representation of the underlying data. If histogram equaliza‐
tion is able to make objects of interest more distinguishable from other objects or
backgrounds (which is not always the case), then it can be a valuable addition to our
image preprocessing pipeline.
8.8 Isolating Colors
Problem
You want to isolate a color in an image.
Solution
Define a range of colors and then apply a mask to the image:
# Load libraries
import cv2
import numpy as np
8.8 Isolating Colors
|
135
from matplotlib import pyplot as plt
# Load image
image_bgr = cv2.imread('images/plane_256x256.jpg')
# Convert BGR to HSV
image_hsv = cv2.cvtColor(image_bgr, cv2.COLOR_BGR2HSV)
# Define range of blue values in HSV
lower_blue = np.array([50,100,50])
upper_blue = np.array([130,255,255])
# Create mask
mask = cv2.inRange(image_hsv, lower_blue, upper_blue)
# Mask image
image_bgr_masked = cv2.bitwise_and(image_bgr, image_bgr, mask=mask)
# Convert BGR to RGB
image_rgb = cv2.cvtColor(image_bgr_masked, cv2.COLOR_BGR2RGB)
# Show image
plt.imshow(image_rgb), plt.axis("off")
plt.show()
136
|
Chapter 8: Handling Images
Discussion
Isolating colors in OpenCV is straightforward. First we convert an image into HSV
(hue, saturation, and value). Second, we define a range of values we want to isolate,
which is probably the most difficult and time-consuming part. Third, we create a
mask for the image (we will only keep the white areas):
# Show image
plt.imshow(mask, cmap='gray'), plt.axis("off")
plt.show()
Finally, we apply the mask to the image using bitwise_and and convert to our
desired output format.
8.9 Binarizing Images
Problem
Given an image, you want to output a simplified version.
8.9 Binarizing Images
|
137
Solution
Thresholding is the process of setting pixels with intensity greater than some value to
be white and less than the value to be black. A more advanced technique is adaptive
thresholding, where the threshold value for a pixel is determined by the pixel intensi‐
ties of its neighbors. This can be helpful when lighting conditions change over differ‐
ent regions in an image:
# Load libraries
import cv2
import numpy as np
from matplotlib import pyplot as plt
# Load image as grayscale
image_grey = cv2.imread("images/plane_256x256.jpg", cv2.IMREAD_GRAYSCALE)
# Apply adaptive thresholding
max_output_value = 255
neighborhood_size = 99
subtract_from_mean = 10
image_binarized = cv2.adaptiveThreshold(image_grey,
max_output_value,
cv2.ADAPTIVE_THRESH_GAUSSIAN_C,
cv2.THRESH_BINARY,
neighborhood_size,
subtract_from_mean)
# Show image
plt.imshow(image_binarized, cmap="gray"), plt.axis("off")
plt.show()
138
|
Chapter 8: Handling Images
Discussion
Our solution has four important arguments in adaptiveThreshold. max_out
put_value simply determines the maximum intensity of the output pixel intensities.
cv2.ADAPTIVE_THRESH_GAUSSIAN_C sets a pixel’s threshold to be a weighted sum of
the neighboring pixel intensities. The weights are determined by a Gaussian window.
Alternatively we could set the threshold to simply the mean of the neighboring pixels
with cv2.ADAPTIVE_THRESH_MEAN_C:
# Apply cv2.ADAPTIVE_THRESH_MEAN_C
image_mean_threshold = cv2.adaptiveThreshold(image_grey,
max_output_value,
cv2.ADAPTIVE_THRESH_MEAN_C,
cv2.THRESH_BINARY,
neighborhood_size,
subtract_from_mean)
# Show image
plt.imshow(image_mean_threshold, cmap="gray"), plt.axis("off")
plt.show()
8.9 Binarizing Images
|
139
The last two parameters are the block size (the size of the neighborhood used to
determine a pixel’s threshold) and a constant subtracted from the calculated thres‐
hold (used to manually fine-tune the threshold).
A major benefit of thresholding is denoising an image—keeping only the most impor‐
tant elements. For example, thresholding is often applied to photos of printed text to
isolate the letters from the page.
8.10 Removing Backgrounds
Problem
You want to isolate the foreground of an image.
Solution
Mark a rectangle around the desired foreground, then run the GrabCut algorithm:
# Load library
import cv2
import numpy as np
from matplotlib import pyplot as plt
140
|
Chapter 8: Handling Images
# Load image and convert to RGB
image_bgr = cv2.imread('images/plane_256x256.jpg')
image_rgb = cv2.cvtColor(image_bgr, cv2.COLOR_BGR2RGB)
# Rectangle values: start x, start y, width, height
rectangle = (0, 56, 256, 150)
# Create initial mask
mask = np.zeros(image_rgb.shape[:2], np.uint8)
# Create temporary arrays used by grabCut
bgdModel = np.zeros((1, 65), np.float64)
fgdModel = np.zeros((1, 65), np.float64)
# Run grabCut
cv2.grabCut(image_rgb, # Our image
mask, # The Mask
rectangle, # Our rectangle
bgdModel, # Temporary array for background
fgdModel, # Temporary array for background
5, # Number of iterations
cv2.GC_INIT_WITH_RECT) # Initiative using our rectangle
# Create mask where sure and likely backgrounds set to 0, otherwise 1
mask_2 = np.where((mask==2) | (mask==0), 0, 1).astype('uint8')
# Multiply image with new mask to subtract background
image_rgb_nobg = image_rgb * mask_2[:, :, np.newaxis]
# Show image
plt.imshow(image_rgb_nobg), plt.axis("off")
plt.show()
8.10 Removing Backgrounds
|
141
Discussion
The first thing we notice is that even though GrabCut did a pretty good job, there are
still areas of background left in the image. We could go back and manually mark
those areas as background, but in the real world we have thousands of images and
manually fixing them individually is not feasible. Therefore, we would do well by
simply accepting that the image data will still contain some background noise.
In our solution, we start out by marking a rectangle around the area that contains the
foreground. GrabCut assumes everything outside this rectangle to be background
and uses that information to figure out what is likely background inside the square
(to learn how the algorithm does this, check out the external resources at the end of
this solution). Then a mask is created that denotes the different definitely/likely back‐
ground/foreground regions:
# Show mask
plt.imshow(mask, cmap='gray'), plt.axis("off")
plt.show()
142
|
Chapter 8: Handling Images
The black region is the area outside our rectangle that is assumed to be definitely
background. The gray area is what GrabCut considered likely background, while the
white area is likely foreground.
This mask is then used to create a second mask that merges the black and gray
regions:
# Show mask
plt.imshow(mask_2, cmap='gray'), plt.axis("off")
plt.show()
8.10 Removing Backgrounds
|
143
The second mask is then applied to the image so that only the foreground remains.
8.11 Detecting Edges
Problem
You want to find the edges in an image.
Solution
Use an edge detection technique like the Canny edge detector:
# Load library
import cv2
import numpy as np
from matplotlib import pyplot as plt
# Load image as grayscale
image_gray = cv2.imread("images/plane_256x256.jpg", cv2.IMREAD_GRAYSCALE)
# Calculate median intensity
median_intensity = np.median(image_gray)
# Set thresholds to be one standard deviation above and below median intensity
144
|
Chapter 8: Handling Images
lower_threshold = int(max(0, (1.0 - 0.33) * median_intensity))
upper_threshold = int(min(255, (1.0 + 0.33) * median_intensity))
# Apply canny edge detector
image_canny = cv2.Canny(image_gray, lower_threshold, upper_threshold)
# Show image
plt.imshow(image_canny, cmap="gray"), plt.axis("off")
plt.show()
Discussion
Edge detection is a major topic of interest in computer vision. Edges are important
because they are areas of high information. For example, in our image one patch of
sky looks very much like another and is unlikely to contain unique or interesting
information. However, patches where the background sky meets the airplane contain
a lot of information (e.g., an object’s shape). Edge detection allows us to remove lowinformation areas and isolate the areas of images containing the most information.
There are many edge detection techniques (Sobel filters, Laplacian edge detector,
etc.). However, our solution uses the commonly used Canny edge detector. How the
Canny detector works is too detailed for this book, but there is one point that we need
to address. The Canny detector requires two parameters denoting low and high gra‐
dient threshold values. Potential edge pixels between the low and high thresholds are
8.11 Detecting Edges
|
145
considered weak edge pixels, while those above the high threshold are considered
strong edge pixels. OpenCV’s Canny method includes the low and high thresholds as
required parameters. In our solution, we set the lower and upper thresholds to be one
standard deviation below and above the image’s median pixel intensity. However,
there are often cases when we might get better results if we used a good pair of low
and high threshold values through manual trial and error using a few images before
running Canny on our entire collection of images.
See Also
• Canny Edge Detector
• Canny Edge Detection Auto Thresholding
8.12 Detecting Corners
Problem
You want to detect the corners in an image.
Solution
Use OpenCV’s implementation of the Harris corner detector, cornerHarris:
# Load libraries
import cv2
import numpy as np
from matplotlib import pyplot as plt
# Load image as grayscale
image_bgr = cv2.imread("images/plane_256x256.jpg")
image_gray = cv2.cvtColor(image_bgr, cv2.COLOR_BGR2GRAY)
image_gray = np.float32(image_gray)
# Set corner detector parameters
block_size = 2
aperture = 29
free_parameter = 0.04
# Detect corners
detector_responses = cv2.cornerHarris(image_gray,
block_size,
aperture,
free_parameter)
# Large corner markers
detector_responses = cv2.dilate(detector_responses, None)
146
|
Chapter 8: Handling Images
# Only keep detector responses greater than threshold, mark as white
threshold = 0.02
image_bgr[detector_responses >
threshold *
detector_responses.max()] = [255,255,255]
# Convert to grayscale
image_gray = cv2.cvtColor(image_bgr, cv2.COLOR_BGR2GRAY)
# Show image
plt.imshow(image_gray, cmap="gray"), plt.axis("off")
plt.show()
Discussion
The Harris corner detector is a commonly used method of detecting the intersection
of two edges. Our interest in detecting corners is motivated by the same reason as for
deleting edges: corners are points of high information. A complete explanation of the
Harris corner detector is available in the external resources at the end of this recipe,
but a simplified explanation is that it looks for windows (also called neighborhoods or
patches) where small movements of the window (imagine shaking the window) cre‐
ates big changes in the contents of the pixels inside the window. cornerHarris con‐
tains three important parameters that we can use to control the edges detected. First,
block_size is the size of the neighbor around each pixel used for corner detection.
8.12 Detecting Corners
|
147
Second, aperture is the size of the Sobel kernel used (don’t worry if you don’t know
what that is), and finally there is a free parameter where larger values correspond to
identifying softer corners.
The output is a grayscale image depicting potential corners:
# Show potential corners
plt.imshow(detector_responses, cmap='gray'), plt.axis("off")
plt.show()
We then apply thresholding to keep only the most likely corners. Alternatively, we
can use a similar detector, the Shi-Tomasi corner detector, which works in a similar
way to the Harris detector (goodFeaturesToTrack) to identify a fixed number of
strong corners. goodFeaturesToTrack takes three major parameters—the number of
corners to detect, the minimum quality of the corner (0 to 1), and the minimum
Euclidean distance between corners:
# Load images
image_bgr = cv2.imread('images/plane_256x256.jpg')
image_gray = cv2.cvtColor(image_bgr, cv2.COLOR_BGR2GRAY)
# Number of corners to detect
corners_to_detect = 10
minimum_quality_score = 0.05
minimum_distance = 25
148
|
Chapter 8: Handling Images
# Detect corners
corners = cv2.goodFeaturesToTrack(image_gray,
corners_to_detect,
minimum_quality_score,
minimum_distance)
corners = np.float32(corners)
# Draw white circle at each corner
for corner in corners:
x, y = corner[0]
cv2.circle(image_bgr, (x,y), 10, (255,255,255), -1)
# Convert to grayscale
image_rgb = cv2.cvtColor(image_bgr, cv2.COLOR_BGR2GRAY)
# Show image
plt.imshow(image_rgb, cmap='gray'), plt.axis("off")
plt.show()
See Also
• OpenCV’s cornerHarris
• OpenCV’s goodFeaturesToTrack
8.12 Detecting Corners
|
149
8.13 Creating Features for Machine Learning
Problem
You want to convert an image into an observation for machine learning.
Solution
Use NumPy’s flatten to convert the multidimensional array containing an image’s
data into a vector containing the observation’s values:
# Load image
import cv2
import numpy as np
from matplotlib import pyplot as plt
# Load image as grayscale
image = cv2.imread("images/plane_256x256.jpg", cv2.IMREAD_GRAYSCALE)
# Resize image to 10 pixels by 10 pixels
image_10x10 = cv2.resize(image, (10, 10))
# Convert image data to one-dimensional vector
image_10x10.flatten()
array([133,
131,
129,
143,
32,
135,
169,
139,
130,
130,
128,
141,
36,
151,
174,
136,
130,
130,
130,
36,
29,
175,
155,
135,
129,
129,
133,
54,
43,
174,
141,
137,
130,
128,
132,
91,
113,
184,
135,
143,
129,
128,
158,
9,
141,
143,
137,
159,
129,
128,
130,
9,
179,
151,
137,
166,
128,
134,
133,
49,
187,
38,
152,
171,
128, 127, 135, 131,
132, 131, 131, 130,
130, 46, 97, 26,
144, 179, 41, 142,
141, 124, 26, 25,
133, 134, 139, 174,
169, 168, 168, 179,
175], dtype=uint8)
131,
129,
132,
95,
132,
177,
152,
Discussion
Images are presented as a grid of pixels. If an image is in grayscale, each pixel is pre‐
sented by one value (i.e., pixel intensity: 1 if white, 0 if black). For example, imagine
we have a 10 × 10–pixel image:
plt.imshow(image_10x10, cmap="gray"), plt.axis("off")
plt.show()
150
|
Chapter 8: Handling Images
In this case the dimensions of the images data will be 10 × 10:
image_10x10.shape
(10, 10)
And if we flatten the array, we get a vector of length 100 (10 multiplied by 10):
image_10x10.flatten().shape
(100,)
This is the feature data for our image that can be joined with the vectors from other
images to create the data we will feed to our machine learning algorithms.
If the image is in color, instead of each pixel being represented by one value, it is rep‐
resented by multiple values (most often three) representing the channels (red, green,
blue, etc.) that blend to make the final color of that pixel. For this reason, if our 10 ×
10 image is in color, we will have 300 feature values for each observation:
# Load image in color
image_color = cv2.imread("images/plane_256x256.jpg", cv2.IMREAD_COLOR)
# Resize image to 10 pixels by 10 pixels
image_color_10x10 = cv2.resize(image_color, (10, 10))
8.13 Creating Features for Machine Learning
|
151
# Convert image data to one-dimensional vector, show dimensions
image_color_10x10.flatten().shape
(300,)
One of the major challenges of image processing and computer vision is that since
every pixel location in a collection of images is a feature, as the images get larger, the
number of features explodes:
# Load image in grayscale
image_256x256_gray = cv2.imread("images/plane_256x256.jpg", cv2.IMREAD_GRAYSCALE)
# Convert image data to one-dimensional vector, show dimensions
image_256x256_gray.flatten().shape
(65536,)
And the number of features only intensifies when the image is in color:
# Load image in color
image_256x256_color = cv2.imread("images/plane_256x256.jpg", cv2.IMREAD_COLOR)
# Convert image data to one-dimensional vector, show dimensions
image_256x256_color.flatten().shape
(196608,)
As the output shows, even a small color image has almost 200,000 features, which can
cause problems when we are training our models because the number of features
might far exceed the number of observations.
This problem will motivate dimensionality strategies discussed in a later chapter,
which attempt to reduce the number of features while not losing an excessive amount
of information contained in the data.
8.14 Encoding Mean Color as a Feature
Problem
You want a feature based on the colors of an image.
Solution
Each pixel in an image is represented by the combination of multiple color channels
(often three: red, green, and blue). Calculate the mean red, green, and blue channel
values for an image to make three color features representing the average colors in
that image:
# Load libraries
import cv2
import numpy as np
152
|
Chapter 8: Handling Images
from matplotlib import pyplot as plt
# Load image as BGR
image_bgr = cv2.imread("images/plane_256x256.jpg", cv2.IMREAD_COLOR)
# Calculate the mean of each channel
channels = cv2.mean(image_bgr)
# Swap blue and red values (making it RGB, not BGR)
observation = np.array([(channels[2], channels[1], channels[0])])
# Show mean channel values
observation
array([[
90.53204346,
133.11735535,
169.03074646]])
We can view the mean channel values directly (apologies to printed book readers):
# Show image
plt.imshow(observation), plt.axis("off")
plt.show()
Discussion
The output is three feature values for an observation, one for each color channel in
the image. These features can be used like any other features in learning algorithms to
classify images according to their colors.
8.15 Encoding Color Histograms as Features
Problem
You want to create a set of features representing the colors appearing in an image.
8.15 Encoding Color Histograms as Features
|
153
Solution
Compute the histograms for each color channel:
# Load libraries
import cv2
import numpy as np
from matplotlib import pyplot as plt
# Load image
image_bgr = cv2.imread("images/plane_256x256.jpg", cv2.IMREAD_COLOR)
# Convert to RGB
image_rgb = cv2.cvtColor(image_bgr, cv2.COLOR_BGR2RGB)
# Create a list for feature values
features = []
# Calculate the histogram for each color channel
colors = ("r","g","b")
# For each channel: calculate histogram and add to feature value list
for i, channel in enumerate(colors):
histogram = cv2.calcHist([image_rgb], # Image
[i], # Index of channel
None, # No mask
[256], # Histogram size
[0,256]) # Range
features.extend(histogram)
# Create a vector for an observation's feature values
observation = np.array(features).flatten()
# Show the observation's value for the first five features
observation[0:5]
array([ 1008.,
217.,
184.,
165.,
116.], dtype=float32)
Discussion
In the RGB color model, each color is the combination of three color channels (i.e.,
red, green, blue). In turn, each channel can take on one of 256 values (represented by
an integer between 0 and 255). For example, the top-leftmost pixel in our image has
the following channel values:
# Show RGB channel values
image_rgb[0,0]
array([107, 163, 212], dtype=uint8)
A histogram is a representation of the distribution of values in data. Here is a simple
example:
154
|
Chapter 8: Handling Images
# Import pandas
import pandas as pd
# Create some data
data = pd.Series([1, 1, 2, 2, 3, 3, 3, 4, 5])
# Show the histogram
data.hist(grid=False)
plt.show()
In this example, we have some data with two 1s, two 2s, three 3s, one 4, and one 5. In
the histogram, each bar represents the number of times each value (1, 2, etc.) appears
in our data.
We can apply this same technique to each of the color channels, but instead of five
possible values, we have 256 (the range of possible values for a channel value). The xaxis represents the 256 possible channel values, and the y-axis represents the number
of times a particular channel value appears across all pixels in an image:
# Calculate the histogram for each color channel
colors = ("r","g","b")
# For each channel: calculate histogram, make plot
for i, channel in enumerate(colors):
histogram = cv2.calcHist([image_rgb], # Image
[i], # Index of channel
None, # No mask
8.15 Encoding Color Histograms as Features
|
155
[256], # Histogram size
[0,256]) # Range
plt.plot(histogram, color = channel)
plt.xlim([0,256])
# Show plot
plt.show()
As we can see in the histogram, barely any pixels contain the blue channel values
between 0 and ~180, while many pixels contain blue channel values between ~190
and ~210. This distribution of channel values is shown for all three channels. The his‐
togram, however, is not simply a visualization; it is 256 features for each color chan‐
nel, making for 768 total features representing the distribution of colors in an image.
See Also
• Histogram
• pandas Histogram documentation
• OpenCV Histogram tutorial
156
|
Chapter 8: Handling Images
CHAPTER 9
Dimensionality Reduction
Using Feature Extraction
9.0 Introduction
It is common to have access to thousands and even hundreds of thousands of fea‐
tures. For example, in Chapter 8 we transformed a 256 × 256–pixel color image into
196,608 features. Furthermore, because each of these pixels can take one of 256 possi‐
ble values, there ends up being 256196608 different configurations our observation can
take. This is problematic because we will practically never be able to collect enough
observations to cover even a small fraction of those configurations and our learning
algorithms do not have enough data to operate correctly.
Fortunately, not all features are created equal and the goal of feature extraction for
dimensionality reduction is to transform our set of features, poriginal, such that we end
up with a new set, pnew, where poriginal > pnew, while still keeping much of the underlying
information. Put another way, we reduce the number of features with only a small
loss in our data’s ability to generate high-quality predictions. In this chapter, we will
cover a number of feature extraction techniques to do just this.
One downside of the feature extraction techniques we discuss is that the new features
we generate will not be interpretable by humans. They will contain as much or nearly
as much ability to train our models, but will appear to the human eye as a collection
of random numbers. If we wanted to maintain our ability to interpret our models,
dimensionality reduction through feature selection is a better option.
157
9.1 Reducing Features Using Principal Components
Problem
Given a set of features, you want to reduce the number of features while retaining the
variance in the data.
Solution
Use principal component analysis with scikit’s PCA:
# Load libraries
from sklearn.preprocessing import StandardScaler
from sklearn.decomposition import PCA
from sklearn import datasets
# Load the data
digits = datasets.load_digits()
# Standardize the feature matrix
features = StandardScaler().fit_transform(digits.data)
# Create a PCA that will retain 99% of variance
pca = PCA(n_components=0.99, whiten=True)
# Conduct PCA
features_pca = pca.fit_transform(features)
# Show results
print("Original number of features:", features.shape[1])
print("Reduced number of features:", features_pca.shape[1])
Original number of features: 64
Reduced number of features: 54
Discussion
Principal component analysis (PCA) is a popular linear dimensionality reduction
technique. PCA projects observations onto the (hopefully fewer) principal compo‐
nents of the feature matrix that retain the most variance. PCA is an unsupervised
technique, meaning that it does not use the information from the target vector and
instead only considers the feature matrix.
For a mathematical description of how PCA works, see the external resources listed at
the end of this recipe. However, we can understand the intuition behind PCA using a
simple example. In the following figure, our data contains two features, x1 and x2.
Looking at the visualization, it should be clear that observations are spread out like a
cigar, with a lot of length and very little height. More specifically, we can say that the
variance of the “length” is significantly greater than the “height.” Instead of length and
158
|
Chapter 9: Dimensionality Reduction Using Feature Extraction
height, we refer to the “directions” with the most variance as the first principal com‐
ponent and the “direction” with the second-most variance as the second principal
component (and so on).
If we wanted to reduce our features, one strategy would be to project all observations
in our 2D space onto the 1D principal component. We would lose the information
captured in the second principal component, but in some situations that would be an
acceptable trade-off. This is PCA.
PCA is implemented in scikit-learn using the pca method. n_components has two
operations, depending on the argument provided. If the argument is greater than 1,
n_components will return that many features. This leads to the question of how to
select the number of features that is optimal. Fortunately for us, if the argument to
n_components is between 0 and 1, pca returns the minimum amount of features that
retain that much variance. It is common to use values of 0.95 and 0.99, meaning 95%
and 99% of the variance of the original features has been retained, respectively.
whiten=True transforms the values of each principal component so that they have
zero mean and unit variance. Another parameter and argument is
svd_solver="randomized", which implements a stochastic algorithm to find the first
principal components in often significantly less time.
The output of our solution shows that PCA let us reduce our dimensionality by 10
features while still retaining 99% of the information (variance) in the feature matrix.
9.1 Reducing Features Using Principal Components
|
159
See Also
• scikit-learn documentation on PCA
• Choosing the Number of Principal Components
• Principal component analysis with linear algebra
9.2 Reducing Features When Data Is Linearly Inseparable
Problem
You suspect you have linearly inseparable data and want to reduce the dimensions.
Solution
Use an extension of principal component analysis that uses kernels to allow for nonlinear dimensionality reduction:
# Load libraries
from sklearn.decomposition import PCA, KernelPCA
from sklearn.datasets import make_circles
# Create linearly inseparable data
features, _ = make_circles(n_samples=1000, random_state=1, noise=0.1, factor=0.1)
# Apply kernal PCA with radius basis function (RBF) kernel
kpca = KernelPCA(kernel="rbf", gamma=15, n_components=1)
features_kpca = kpca.fit_transform(features)
print("Original number of features:", features.shape[1])
print("Reduced number of features:", features_kpca.shape[1])
Original number of features: 2
Reduced number of features: 1
Discussion
PCA is able to reduce the dimensionality of our feature matrix (e.g., the number of
features). Standard PCA uses linear projection to reduce the features. If the data is
linearly separable (i.e., you can draw a straight line or hyperplane between different
classes) then PCA works well. However, if your data is not linearly separable (e.g., you
can only separate classes using a curved decision boundary), the linear transforma‐
tion will not work as well. In our solution we used scikit-learn’s make_circles to gen‐
erate a simulated dataset with a target vector of two classes and two features.
make_circles makes linearly inseparable data; specifically, one class is surrounded
on all sides by the other class.
160
|
Chapter 9: Dimensionality Reduction Using Feature Extraction
If we used linear PCA to reduce the dimensions of our data, the two classes would be
linearly projected onto the first principal component such that they would become
intertwined.
Ideally, we would want a transformation that would both reduce the dimensions and
also make the data linearly separable. Kernel PCA can do both.
Kernels allow us to project the linearly inseparable data into a higher dimension
where it is linearly separable; this is called the kernel trick. Don’t worry if you don’t
understand the details of the kernel trick; just think of kernels as different ways of
projecting the data. There are a number of kernels we can use in scikit-learn’s ker
nelPCA, specified using the kernel parameter. A common kernel to use is the Gaus‐
sian radial basis function kernel rbf, but other options are the polynomial kernel
(poly) and sigmoid kernel (sigmoid). We can even specify a linear projection (lin
ear), which will produce the same results as standard PCA.
9.2 Reducing Features When Data Is Linearly Inseparable
|
161
One downside of kernel PCA is that there are a number of parameters we need to
specify. For example, in Recipe 9.1 we set n_components to 0.99 to make PCA select
the number of components to retain 99% of the variance. We don’t have this option in
kernel PCA. Instead we have to define the number of parameters (e.g., n_compo
nents=1). Furthermore, kernels come with their own hyperparameters that we will
have to set; for example, the radial basis function requires a gamma value.
So how do we know which values to use? Through trial and error. Specifically we can
train our machine learning model multiple times, each time with a different kernel or
different value of the parameter. Once we find the combination of values that pro‐
duces the highest quality predicted values, we are done. We will learn about this strat‐
egy in depth in Chapter 12.
See Also
• scikit-learn documentation on Kernel PCA
• Kernel tricks and nonlinear dimensionality reduction via RBF kernel PCA
9.3 Reducing Features by Maximizing Class Separability
Problem
You want to reduce the features to be used by a classifier.
Solution
Try linear discriminant analysis (LDA) to project the features onto component axes
that maximize the separation of classes:
# Load libraries
from sklearn import datasets
from sklearn.discriminant_analysis import LinearDiscriminantAnalysis
# Load Iris flower dataset:
iris = datasets.load_iris()
features = iris.data
target = iris.target
# Create and run an LDA, then use it to transform the features
lda = LinearDiscriminantAnalysis(n_components=1)
features_lda = lda.fit(features, target).transform(features)
# Print the number of features
print("Original number of features:", features.shape[1])
print("Reduced number of features:", features_lda.shape[1])
162
|
Chapter 9: Dimensionality Reduction Using Feature Extraction
Original number of features: 4
Reduced number of features: 1
We can use explained_variance_ratio_ to view the amount of variance explained
by each component. In our solution the single component explained over 99% of the
variance:
lda.explained_variance_ratio_
array([ 0.99147248])
Discussion
LDA is a classification that is also a popular technique for dimensionality reduction.
LDA works similarly to principal component analysis (PCA) in that it projects our
feature space onto a lower-dimensional space. However, in PCA we were only interes‐
ted in the component axes that maximize the variance in the data, while in LDA we
have the additional goal of maximizing the differences between classes. In this pic‐
tured example, we have data comprising two target classes and two features. If we
project the data onto the y-axis, the two classes are not easily separable (i.e., they
overlap), while if we project the data onto the x-axis, we are left with a feature vector
(i.e., we reduced our dimensionality by one) that still preserves class separability. In
the real world, of course, the relationship between the classes will be more complex
and the dimensionality will be higher, but the concept remains the same.
9.3 Reducing Features by Maximizing Class Separability
|
163
In scikit-learn, LDA is implemented using LinearDiscriminantAnalysis, which
includes a parameter, n_components, indicating the number of features we want
returned. To figure out what argument value to use with n_components (e.g., how
many parameters to keep), we can take advantage of the fact that explained_var
iance_ratio_ tells us the variance explained by each outputted feature and is a sor‐
ted array. For example:
lda.explained_variance_ratio_
array([ 0.99147248])
Specifically, we can run LinearDiscriminantAnalysis with n_components set to
None to return the ratio of variance explained by every component feature, then cal‐
culate how many components are required to get above some threshold of variance
explained (often 0.95 or 0.99):
# Create and run LDA
lda = LinearDiscriminantAnalysis(n_components=None)
features_lda = lda.fit(features, target)
# Create array of explained variance ratios
lda_var_ratios = lda.explained_variance_ratio_
# Create function
def select_n_components(var_ratio, goal_var: float) -> int:
# Set initial variance explained so far
total_variance = 0.0
# Set initial number of features
n_components = 0
# For the explained variance of each feature:
for explained_variance in var_ratio:
# Add the explained variance to the total
total_variance += explained_variance
# Add one to the number of components
n_components += 1
# If we reach our goal level of explained variance
if total_variance >= goal_var:
# End the loop
break
# Return the number of components
return n_components
# Run function
select_n_components(lda_var_ratios, 0.95)
1
164
|
Chapter 9: Dimensionality Reduction Using Feature Extraction
See Also
• Comparison of LDA and PCA 2D projection of Iris dataset
• Linear Discriminant Analysis
9.4 Reducing Features Using Matrix Factorization
Problem
You have a feature matrix of nonnegative values and want to reduce the dimensional‐
ity.
Solution
Use non-negative matrix factorization (NMF) to reduce the dimensionality of the fea‐
ture matrix:
# Load libraries
from sklearn.decomposition import NMF
from sklearn import datasets
# Load the data
digits = datasets.load_digits()
# Load feature matrix
features = digits.data
# Create, fit, and apply NMF
nmf = NMF(n_components=10, random_state=1)
features_nmf = nmf.fit_transform(features)
# Show results
print("Original number of features:", features.shape[1])
print("Reduced number of features:", features_nmf.shape[1])
Original number of features: 64
Reduced number of features: 10
Discussion
NMF is an unsupervised technique for linear dimensionality reduction that factorizes
(i.e., breaks up into multiple matrices whose product approximates the original
matrix) the feature matrix into matrices representing the latent relationship between
observations and their features. Intuitively, NMF can reduce dimensionality because
in matrix multiplication, the two factors (matrices being multiplied) can have signifi‐
cantly fewer dimensions than the product matrix. Formally, given a desired number
of returned features, r, NMF factorizes our feature matrix such that:
9.4 Reducing Features Using Matrix Factorization
|
165
V ≈ WH
where V is our d × _n feature matrix (i.e., d features, n observations), W is a d × r,
and H is an r × n matrix. By adjusting the value of r we can set the amount of dimen‐
sionality reduction desired.
One major requirement of NMA is that, as the name implies, the feature matrix can‐
not contain negative values. Additionally, unlike PCA and other techniques we have
examined, NMA does not provide us with the explained variance of the outputted
features. Thus, the best way for us to find the optimum value of n_components is by
trying a range of values to find the one that produces the best result in our end model
(see Chapter 12).
See Also
• Non-Negative Matrix Factorization (NMF)
9.5 Reducing Features on Sparse Data
Problem
You have a sparse feature matrix and want to reduce the dimensionality.
Solution
Use Truncated Singular Value Decomposition (TSVD):
# Load libraries
from sklearn.preprocessing import StandardScaler
from sklearn.decomposition import TruncatedSVD
from scipy.sparse import csr_matrix
from sklearn import datasets
import numpy as np
# Load the data
digits = datasets.load_digits()
# Standardize feature matrix
features = StandardScaler().fit_transform(digits.data)
# Make sparse matrix
features_sparse = csr_matrix(features)
# Create a TSVD
tsvd = TruncatedSVD(n_components=10)
# Conduct TSVD on sparse matrix
166
|
Chapter 9: Dimensionality Reduction Using Feature Extraction
features_sparse_tsvd = tsvd.fit(features_sparse).transform(features_sparse)
# Show results
print("Original number of features:", features_sparse.shape[1])
print('"Reduced number of features:", features_sparse_tsvd.shape[1])
Original number of features: 64
Reduced number of features: 10
Discussion
TSVD is similar to PCA and in fact, PCA actually often uses non-truncated Singular
Value Decomposition (SVD) in one of its steps. In regular SVD, given d features,
SVD will create factor matrices that are d × d, whereas TSVD will return factors that
are n × n, where n is previously specified by a parameter. The practical advantage of
TSVD is that unlike PCA, it works on sparse feature matrices.
One issue with TSVD is that because of how it uses a random number generator, the
signs of the output can flip between fittings. An easy workaround is to use fit only
once per preprocessing pipeline, then use transform multiple times.
As with linear discriminant analysis, we have to specify the number of features (com‐
ponents) we want outputted. This is done with the n_components parameter. A natu‐
ral question is then: what is the optimum number of components? One strategy is to
include n_components as a hyperparameter to optimize during model selection (i.e.,
choose the value for n_components that produces the best trained model). Alterna‐
tively, because TSVD provides us with the ratio of the original feature matrix’s var‐
iance explained by each component, we can select the number of components that
explain a desired amount of variance (95% or 99% are common values). For example,
in our solution the first three outputted components explain approximately 30% of
the original data’s variance:
# Sum of first three components' explained variance ratios
tsvd.explained_variance_ratio_[0:3].sum()
0.30039385386597783
We can automate the process by creating a function that runs TSVD with n_compo
nents set to one less than the number of original features and then calculate the num‐
ber of components that explain a desired amount of the original data’s variance:
# Create and run an TSVD with one less than number of features
tsvd = TruncatedSVD(n_components=features_sparse.shape[1]-1)
features_tsvd = tsvd.fit(features)
# List of explained variances
tsvd_var_ratios = tsvd.explained_variance_ratio_
# Create a function
def select_n_components(var_ratio, goal_var):
9.5 Reducing Features on Sparse Data
|
167
# Set initial variance explained so far
total_variance = 0.0
# Set initial number of features
n_components = 0
# For the explained variance of each feature:
for explained_variance in var_ratio:
# Add the explained variance to the total
total_variance += explained_variance
# Add one to the number of components
n_components += 1
# If we reach our goal level of explained variance
if total_variance >= goal_var:
# End the loop
break
# Return the number of components
return n_components
# Run function
select_n_components(tsvd_var_ratios, 0.95)
40
See Also
• scikit-learn documentation TruncatedSVD
168
|
Chapter 9: Dimensionality Reduction Using Feature Extraction
CHAPTER 10
Dimensionality Reduction
Using Feature Selection
10.0 Introduction
In Chapter 9, we discussed how to reduce the dimensionality of our feature matrix by
creating new features with (ideally) similar ability to train quality models but with
significantly fewer dimensions. This is called feature extraction. In this chapter we
will cover an alternative approach: selecting high-quality, informative features and
dropping less useful features. This is called feature selection.
There are three types of feature selection methods: filter, wrapper, and embedded. Fil‐
ter methods select the best features by examining their statistical properties. Wrapper
methods use trial and error to find the subset of features that produce models with
the highest quality predictions. Finally, embedded methods select the best feature
subset as part or as an extension of a learning algorithm’s training process.
Ideally, we’d describe all three methods in this chapter. However, since embedded
methods are closely intertwined with specific learning algorithms, they are difficult to
explain prior to a deeper dive into the algorithms themselves. Therefore, in this chap‐
ter we cover only filter and wrapper feature selection methods, leaving the discussion
of particular embedded methods until the chapters where those learning algorithms
are discussed in depth.
169
10.1 Thresholding Numerical Feature Variance
Problem
You have a set of numerical features and want to remove those with low variance (i.e.,
likely containing little information).
Solution
Select a subset of features with variances above a given threshold:
# Load libraries
from sklearn import datasets
from sklearn.feature_selection import VarianceThreshold
# import some data to play with
iris = datasets.load_iris()
# Create features and target
features = iris.data
target = iris.target
# Create thresholder
thresholder = VarianceThreshold(threshold=.5)
# Create high variance feature matrix
features_high_variance = thresholder.fit_transform(features)
# View high variance feature matrix
features_high_variance[0:3]
array([[ 5.1,
[ 4.9,
[ 4.7,
1.4,
1.4,
1.3,
0.2],
0.2],
0.2]])
Discussion
Variance thresholding (VT) is one of the most basic approaches to feature selection. It
is motivated by the idea that features with low variance are likely less interesting (and
useful) than features with high variance. VT first calculates the variance of each
feature:
operatornameVar x =
1 n
xi − μ
ni∑
=1
2
where x is the feature vector, xi is an individual feature value, and μ is that feature’s
mean value. Next, it drops all features whose variance does not meet that threshold.
170
|
Chapter 10: Dimensionality Reduction Using Feature Selection
There are two things to keep in mind when employing VT. First, the variance is not
centered; that is, it is in the squared unit of the feature itself. Therefore, the VT will
not work when feature sets contain different units (e.g., one feature is in years while a
different feature is in dollars). Second, the variance threshold is selected manually, so
we have to use our own judgment for a good value to select (or use a model selection
technique described in Chapter 12). We can see the variance for each feature using
variances_:
# View variances
thresholder.fit(features).variances_
array([ 0.68112222,
0.18675067,
3.09242489,
0.57853156])
Finally, if the features have been standardized (to mean zero and unit variance), then
for obvious reasons variance thresholding will not work correctly:
# Load library
from sklearn.preprocessing import StandardScaler
# Standardize feature matrix
scaler = StandardScaler()
features_std = scaler.fit_transform(features)
# Caculate variance of each feature
selector = VarianceThreshold()
selector.fit(features_std).variances_
array([ 1.,
1.,
1.,
1.])
10.2 Thresholding Binary Feature Variance
Problem
You have a set of binary categorical features and want to remove those with low var‐
iance (i.e., likely containing little information).
Solution
Select a subset of features with a Bernoulli random variable variance above a given
threshold:
# Load library
from sklearn.feature_selection import VarianceThreshold
# Create feature matrix with:
# Feature 0: 80% class 0
# Feature 1: 80% class 1
# Feature 2: 60% class 0, 40% class 1
features = [[0, 1, 0],
[0, 1, 1],
10.2 Thresholding Binary Feature Variance
|
171
[0, 1, 0],
[0, 1, 1],
[1, 0, 0]]
# Run threshold by variance
thresholder = VarianceThreshold(threshold=(.75 * (1 - .75)))
thresholder.fit_transform(features)
array([[0],
[1],
[0],
[1],
[0]])
Discussion
Just like with numerical features, one strategy for selecting highly informative catego‐
rical features is to examine their variances. In binary features (i.e., Bernoulli random
variables), variance is calculated as:
Var x = p 1 − p
where p is the proportion of observations of class 1. Therefore, by setting p, we can
remove features where the vast majority of observations are one class.
10.3 Handling Highly Correlated Features
Problem
You have a feature matrix and suspect some features are highly correlated.
Solution
Use a correlation matrix to check for highly correlated features. If highly correlated
features exist, consider dropping one of the correlated features:
# Load libraries
import pandas as pd
import numpy as np
# Create feature matrix with two highly correlated features
features = np.array([[1, 1, 1],
[2, 2, 0],
[3, 3, 1],
[4, 4, 0],
[5, 5, 1],
[6, 6, 0],
[7, 7, 1],
[8, 7, 0],
172
|
Chapter 10: Dimensionality Reduction Using Feature Selection
[9, 7, 1]])
# Convert feature matrix into DataFrame
dataframe = pd.DataFrame(features)
# Create correlation matrix
corr_matrix = dataframe.corr().abs()
# Select upper triangle of correlation matrix
upper = corr_matrix.where(np.triu(np.ones(corr_matrix.shape),
k=1).astype(np.bool))
# Find index of feature columns with correlation greater than 0.95
to_drop = [column for column in upper.columns if any(upper[column] > 0.95)]
# Drop features
dataframe.drop(dataframe.columns[to_drop], axis=1).head(3)
0 2
0 1 1
1 2 0
2 3 1
Discussion
One problem we often run into in machine learning is highly correlated features. If
two features are highly correlated, then the information they contain is very similar,
and it is likely redundant to include both features. The solution to highly correlated
features is simple: remove one of them from the feature set.
In our solution, first we create a correlation matrix of all features:
# Correlation matrix
dataframe.corr()
0
1
0 1.000000 0.976103
1 0.976103 1.000000
2 0.000000 -0.034503
2
0.000000
-0.034503
1.000000
Second, we look at the upper triangle of the correlation matrix to identify pairs of
highly correlated features:
# Upper triangle of correlation matrix
upper
10.3 Handling Highly Correlated Features
|
173
0
1
0 NaN 0.976103
1 NaN NaN
2 NaN NaN
2
0.000000
0.034503
NaN
Third, we remove one feature from each of those pairs from the feature set.
10.4 Removing Irrelevant Features for Classification
Problem
You have a categorical target vector and want to remove uninformative features.
Solution
If the features are categorical, calculate a chi-square (χ2) statistic between each feature
and the target vector:
# Load libraries
from sklearn.datasets import load_iris
from sklearn.feature_selection import SelectKBest
from sklearn.feature_selection import chi2, f_classif
# Load data
iris = load_iris()
features = iris.data
target = iris.target
# Convert to categorical data by converting data to integers
features = features.astype(int)
# Select two features with highest chi-squared statistics
chi2_selector = SelectKBest(chi2, k=2)
features_kbest = chi2_selector.fit_transform(features, target)
# Show results
print("Original number of features:", features.shape[1])
print("Reduced number of features:", features_kbest.shape[1])
Original number of features: 4
Reduced number of features: 2
If the features are quantitative, compute the ANOVA F-value between each feature
and the target vector:
# Select two features with highest F-values
fvalue_selector = SelectKBest(f_classif, k=2)
features_kbest = fvalue_selector.fit_transform(features, target)
174
|
Chapter 10: Dimensionality Reduction Using Feature Selection
# Show results
print("Original number of features:", features.shape[1])
print("Reduced number of features:", features_kbest.shape[1])
Original number of features: 4
Reduced number of features: 2
Instead of selecting a specific number of features, we can also use SelectPercentile
to select the top n percent of features:
# Load library
from sklearn.feature_selection import SelectPercentile
# Select top 75% of features with highest F-values
fvalue_selector = SelectPercentile(f_classif, percentile=75)
features_kbest = fvalue_selector.fit_transform(features, target)
# Show results
print("Original number of features:", features.shape[1])
print("Reduced number of features:", features_kbest.shape[1])
Original number of features: 4
Reduced number of features: 3
Discussion
Chi-square statistics examines the independence of two categorical vectors. That is,
the statistic is the difference between the observed number of observations in each
class of a categorical feature and what we would expect if that feature was independ‐
ent (i.e., no relationship) with the target vector:
χ2 =
n
∑
i=1
Oi − Ei
Ei
2
where Oi is the number of observations in class i and Ei is the number of observations
in class i we would expect if there is no relationship between the feature and target
vector.
A chi-squared statistic is a single number that tells you how much difference exists
between your observed counts and the counts you would expect if there were no rela‐
tionship at all in the population. By calculating the chi-squared statistic between a
feature and the target vector, we obtain a measurement of the independence between
the two. If the target is independent of the feature variable, then it is irrelevant for our
purposes because it contains no information we can use for classification. On the
other hand, if the two features are highly dependent, they likely are very informative
for training our model.
10.4 Removing Irrelevant Features for Classification |
175
To use chi-squared in feature selection, we calculate the chi-squared statistic between
each feature and the target vector, then select the features with the best chi-square
statistics. In scikit-learn, we can use SelectKBest to select the features with the best
statistics. The parameter k determines the number of features we want to keep.
It is important to note that chi-square statistics can only be calculated between two
categorical vectors. For this reason, chi-squared for feature selection requires that
both the target vector and the features are categorical. However, if we have a numeri‐
cal feature we can use the chi-squared technique by first transforming the quantita‐
tive feature into a categorical feature. Finally, to use our chi-squared approach, all val‐
ues need to be non-negative.
Alternatively, if we have a numerical feature we can use f_classif to calculate the
ANOVA F-value statistic with each feature and the target vector. F-value scores exam‐
ine if, when we group the numerical feature by the target vector, the means for each
group are significantly different. For example, if we had a binary target vector, gender,
and a quantitative feature, test scores, the F-value score would tell us if the mean test
score for men is different than the mean test score for women. If it is not, then test
score doesn’t help us predict gender and therefore the feature is irrelevant.
10.5 Recursively Eliminating Features
Problem
You want to automatically select the best features to keep.
Solution
Use scikit-learn’s RFECV to conduct recursive feature elimination (RFE) using crossvalidation (CV). That is, repeatedly train a model, each time removing a feature until
model performance (e.g., accuracy) becomes worse. The remaining features are the
best:
# Load libraries
import warnings
from sklearn.datasets import make_regression
from sklearn.feature_selection import RFECV
from sklearn import datasets, linear_model
# Suppress an annoying but harmless warning
warnings.filterwarnings(action="ignore", module="scipy",
message="^internal gelsd")
# Generate features matrix, target vector, and the true coefficients
features, target = make_regression(n_samples = 10000,
n_features = 100,
n_informative = 2,
176
|
Chapter 10: Dimensionality Reduction Using Feature Selection
random_state = 1)
# Create a linear regression
ols = linear_model.LinearRegression()
# Recursively eliminate features
rfecv = RFECV(estimator=ols, step=1, scoring="neg_mean_squared_error")
rfecv.fit(features, target)
rfecv.transform(features)
array([[ 0.00850799, 0.7031277 , -1.2416911 , -0.25651883, -0.10738769],
[-1.07500204, 2.56148527, 0.5540926 , -0.72602474, -0.91773159],
[ 1.37940721, -1.77039484, -0.59609275, 0.51485979, -1.17442094],
...,
[-0.80331656, -1.60648007, 0.37195763, 0.78006511, -0.20756972],
[ 0.39508844, -1.34564911, -0.9639982 , 1.7983361 , -0.61308782],
[-0.55383035, 0.82880112, 0.24597833, -1.71411248, 0.3816852 ]])
Once we have conducted RFE, we can see the number of features we should keep:
# Number of best features
rfecv.n_features_
5
We can also see which of those features we should keep:
# Which categories are best
rfecv.support_
array([False, False, False,
False, False, False,
False, False, False,
False, False, False,
False, False, False,
False, False, False,
False, False, False,
False, False, False,
False, False, False,
False, False, False,
False, False, False,
False], dtype=bool)
False,
False,
False,
False,
True,
False,
False,
False,
False,
False,
False,
False,
False,
False,
False,
False,
False,
False,
False,
False,
False,
False,
True,
False,
False,
False,
False,
False,
False,
True,
False,
False,
False,
False,
False,
False,
False,
False,
False,
True,
False,
False,
False,
False,
False,
False,
False,
False,
True,
False,
False,
False,
False,
False,
False,
False,
False,
False,
False,
False,
False,
False,
False,
False,
False,
False,
We can even view the rankings of the features:
# Rank features best (1) to worst
rfecv.ranking_
array([11,
84,
82,
24,
1,
51,
92,
54,
45,
32,
95,
64,
96,
88,
89,
13,
77,
6,
87,
37,
69,
49,
93,
9,
46, 1, 48, 23, 16, 2,
42, 85, 62, 74, 50, 80,
94, 1, 35, 47, 39, 1,
26, 12, 71, 68, 40, 1,
22, 52, 30, 60, 81, 14,
20, 5, 55, 56, 25, 36,
66,
10,
34,
43,
86,
61,
83,
38,
72,
63,
18,
78,
33,
59,
19,
28,
15,
31,
27,
79,
4,
73,
41,
3,
70, 75,
57, 44,
17, 91,
58, 21,
7, 53,
76])
29,
8,
90,
67,
65,
10.5 Recursively Eliminating Features
|
177
Discussion
This is likely the most advanced recipe in this book up to this point, combining a
number of topics we have yet to address in detail. However, the intuition is straight‐
forward enough that we can address it here rather than holding off until a later chap‐
ter. The idea behind RFE is to train a model that contains some parameters (also
called weights or coefficients) like linear regression or support vector machines repeat‐
edly. The first time we train the model, we include all the features. Then, we find the
feature with the smallest parameter (notice that this assumes the features are either
rescaled or standardized), meaning it is less important, and remove the feature from
the feature set.
The obvious question then is: how many features should we keep? We can (hypotheti‐
cally) repeat this loop until we only have one feature left. A better approach requires
that we include a new concept called cross-validation (CV). We will discuss crossvalidation in detail in the next chapter, but here is the general idea.
Given data containing 1) a target we want to predict and 2) a feature matrix, first we
split the data into two groups: a training set and a test set. Second, we train our model
using the training set. Third, we pretend that we do not know the target of the test
set, and apply our model to the test set’s features in order to predict the values of the
test set. Finally, we compare our predicted target values with the true target values to
evaluate our model.
We can use CV to find the optimum number of features to keep during RFE. Specifi‐
cally, in RFE with CV after every iteration, we use cross-validation to evaluate our
model. If CV shows that our model improved after we eliminated a feature, then we
continue on to the next loop. However, if CV shows that our model got worse after
we eliminated a feature, we put that feature back into the feature set and select those
features as the best.
In scikit-learn, RFE with CV is implemented using RFECV and contains a number of
important parameters. The estimator parameter determines the type of model we
want to train (e.g., linear regression). The step regression sets the number or propor‐
tion of features to drop during each loop. The scoring parameter sets the metric of
quality we use to evaluate our model during cross-validation.
See Also
• Recursive feature elimination with cross-validation
178
|
Chapter 10: Dimensionality Reduction Using Feature Selection
CHAPTER 11
Model Evaluation
11.0 Introduction
In this chapter we will examine strategies for evaluating the quality of models created
through our learning algorithms. It might appear strange to discuss model evaluation
before discussing how to create them, but there is a method to our madness. Models
are only as useful as the quality of their predictions, and thus fundamentally our goal
is not to create models (which is easy) but to create high-quality models (which is
hard). Therefore, before we explore the myriad learning algorithms, we first set up
how we can evaluate the models they produce.
11.1 Cross-Validating Models
Problem
You want to evaluate how well your model will work in the real world.
Solution
Create a pipeline that preprocesses the data, trains the model, and then evaluates it
using cross-validation:
# Load libraries
from sklearn import datasets
from sklearn import metrics
from sklearn.model_selection import KFold, cross_val_score
from sklearn.pipeline import make_pipeline
from sklearn.linear_model import LogisticRegression
from sklearn.preprocessing import StandardScaler
# Load digits dataset
179
digits = datasets.load_digits()
# Create features matrix
features = digits.data
# Create target vector
target = digits.target
# Create standardizer
standardizer = StandardScaler()
# Create logistic regression object
logit = LogisticRegression()
# Create a pipeline that standardizes, then runs logistic regression
pipeline = make_pipeline(standardizer, logit)
# Create k-Fold cross-validation
kf = KFold(n_splits=10, shuffle=True, random_state=1)
# Conduct k-fold cross-validation
cv_results = cross_val_score(pipeline, # Pipeline
features, # Feature matrix
target, # Target vector
cv=kf, # Cross-validation technique
scoring="accuracy", # Loss function
n_jobs=-1) # Use all CPU scores
# Calculate mean
cv_results.mean()
0.96493171942892597
Discussion
At first consideration, evaluating supervised-learning models might appear straight‐
forward: train a model and then calculate how well it did using some performance
metric (accuracy, squared errors, etc.). However, this approach is fundamentally
flawed. If we train a model using our data, and then evaluate how well it did on that
data, we are not achieving our desired goal. Our goal is not to evaluate how well the
model does on our training data, but how well it does on data it has never seen before
(e.g., a new customer, a new crime, a new image). For this reason, our method of
evaluation should help us understand how well models are able to make predictions
from data they have never seen before.
One strategy might be to hold off a slice of data for testing. This is called validation
(or hold-out). In validation our observations (features and targets) are split into two
sets, traditionally called the training set and the test set. We take the test set and put it
off to the side, pretending that we have never seen it before. Next we train our model
using our training set, using the features and target vector to teach the model how to
180
|
Chapter 11: Model Evaluation
make the best prediction. Finally, we simulate having never before seen external data
by evaluating how our model trained on our training set performs on our test set.
However, the validation approach has two major weaknesses. First, the performance
of the model can be highly dependent on which few observations were selected for
the test set. Second, the model is not being trained using all the available data, and
not being evaluated on all the available data.
A better strategy, which overcomes these weaknesses, is called k-fold cross-validation
(KFCV). In KFCV, we split the data into k parts called “folds.” The model is then
trained using k – 1 folds—combined into one training set—and then the last fold is
used as a test set. We repeat this k times, each time using a different fold as the test
set. The performance on the model for each of the k iterations is then averaged to
produce an overall measurement.
In our solution, we conducted k-fold cross-validation using 10 folds and outputted
the evaluation scores to cv_results:
# View score for all 10 folds
cv_results
array([ 0.97222222,
0.98333333,
0.97777778,
0.97777778,
0.95555556,
0.96648045,
0.95
,
0.96089385,
0.95555556,
0.94972067])
There are three important points to consider when we are using KFCV. First, KFCV
assumes that each observation was created independent from the other (i.e., the data
is independent identically distributed [IID]). If the data is IID, it is a good idea to
shuffle observations when assigning to folds. In scikit-learn we can set shuffle=True
to perform shuffling.
Second, when we are using KFCV to evaluate a classifier, it is often beneficial to have
folds containing roughly the same percentage of observations from each of the differ‐
ent target classes (called stratified k-fold). For example, if our target vector contained
gender and 80% of the observations were male, then each fold would contain 80%
male and 20% female observations. In scikit-learn, we can conduct stratified k-fold
cross-validation by replacing the KFold class with StratifiedKFold.
Finally, when we are using validation sets or cross-validation, it is important to pre‐
process data based on the training set and then apply those transformations to both
the training and test set. For example, when we fit our standardization object, stand
ardizer, we calculate the mean and variance of only the training set. Then we apply
that transformation (using transform) to both the training and test sets:
# Import library
from sklearn.model_selection import train_test_split
# Create training and test sets
features_train, features_test, target_train, target_test = train_test_split(
features, target, test_size=0.1, random_state=1)
11.1 Cross-Validating Models
|
181
# Fit standardizer to training set
standardizer.fit(features_train)
# Apply to both training and test sets
features_train_std = standardizer.transform(features_train)
features_test_std = standardizer.transform(features_test)
The reason for this is because we are pretending that the test set is unknown data. If
we fit both our preprocessors using observations from both training and test sets,
some of the information from the test set leaks into our training set. This rule applies
for any preprocessing step such as feature selection.
scikit-learn’s pipeline package makes this easy to do while using cross-validation tech‐
niques. We first create a pipeline that preprocesses the data (e.g., standardizer) and
then trains a model (logistic regression, logit):
# Create a pipeline
pipeline = make_pipeline(standardizer, logit)
Then we run KFCV using that pipeline and scikit does all the work for us:
# Do k-fold cross-validation
cv_results = cross_val_score(pipeline, # Pipeline
features, # Feature matrix
target, # Target vector
cv=kf, # Cross-validation technique
scoring="accuracy", # Loss function
n_jobs=-1) # Use all CPU scores
cross_val_score comes with three parameters that we have not discussed that are
worth noting. cv determines our cross-validation technique. K-fold is the most com‐
mon by far, but there are others, like leave-one-out-cross-validation where the num‐
ber of folds k equals the number of observations. The scoring parameter defines our
metric for success, a number of which are discussed in other recipes in this chapter.
Finally, n_jobs=-1 tells scikit-learn to use every core available. For example, if your
computer has four cores (a common number for laptops), then scikit-learn will use
all four cores at once to speed up the operation.
See Also
• Why every statistician should know about cross-validation
• Cross-Validation Gone Wrong
182
|
Chapter 11: Model Evaluation
11.2 Creating a Baseline Regression Model
Problem
You want a simple baseline regression model to compare against your model.
Solution
Use scikit-learn’s DummyRegressor to create a simple model to use as a baseline:
# Load libraries
from sklearn.datasets import load_boston
from sklearn.dummy import DummyRegressor
from sklearn.model_selection import train_test_split
# Load data
boston = load_boston()
# Create features
features, target = boston.data, boston.target
# Make test and training split
features_train, features_test, target_train, target_test = train_test_split(
features, target, random_state=0)
# Create a dummy regressor
dummy = DummyRegressor(strategy='mean')
# "Train" dummy regressor
dummy.fit(features_train, target_train)
# Get R-squared score
dummy.score(features_test, target_test)
-0.0011193592039553391
To compare, we train our model and evaluate the performance score:
# Load library
from sklearn.linear_model import LinearRegression
# Train simple linear regression model
ols = LinearRegression()
ols.fit(features_train, target_train)
# Get R-squared score
ols.score(features_test, target_test)
0.63536207866746675
11.2 Creating a Baseline Regression Model
|
183
Discussion
DummyRegressor allows us to create a very simple model that we can use as a baseline
to compare against our actual model. This can often be useful to simulate a “naive”
existing prediction process in a product or system. For example, a product might
have been originally hardcoded to assume that all new users will spend $100 in the
first month, regardless of their features. If we encode that assumption into a baseline
model, we are able to concretely state the benefits of using a machine learning
approach.
DummyRegressor uses the strategy parameter to set the method of making predic‐
tions, including the mean or median value in the training set. Furthermore, if we set
strategy to constant and use the constant parameter, we can set the dummy
regressor to predict some constant value for every observation:
# Create dummy regressor that predicts 20's for everything
clf = DummyRegressor(strategy='constant', constant=20)
clf.fit(features_train, target_train)
# Evaluate score
clf.score(features_test, target_test)
-0.065105020293257265
One small note regarding score. By default, score returns the coefficient of determi‐
nation (R-squared, R2) score:
R2 = 1 −
∑i yi − y i
∑i yi − y
2
2
where yi is the true value of the target observation, y i is the predicted value, and ȳ is
the mean value for the target vector.
The closer R2 is to 1, the more of the variance in the target vector that is explained by
the features.
11.3 Creating a Baseline Classification Model
Problem
You want a simple baseline classifier to compare against your model.
Solution
Use scikit-learn’s DummyClassifier:
184
|
Chapter 11: Model Evaluation
# Load libraries
from sklearn.datasets import load_iris
from sklearn.dummy import DummyClassifier
from sklearn.model_selection import train_test_split
# Load data
iris = load_iris()
# Create target vector and feature matrix
features, target = iris.data, iris.target
# Split into training and test set
features_train, features_test, target_train, target_test = train_test_split(
features, target, random_state=0)
# Create dummy classifier
dummy = DummyClassifier(strategy='uniform', random_state=1)
# "Train" model
dummy.fit(features_train, target_train)
# Get accuracy score
dummy.score(features_test, target_test)
0.42105263157894735
By comparing the baseline classifier to our trained classifier, we can see the improve‐
ment:
# Load library
from sklearn.ensemble import RandomForestClassifier
# Create classifier
classifier = RandomForestClassifier()
# Train model
classifier.fit(features_train, target_train)
# Get accuracy score
classifier.score(features_test, target_test)
0.94736842105263153
Discussion
A common measure of a classifier’s performance is how much better it is than ran‐
dom guessing. scikit-learn’s DummyClassifier makes this comparison easy. The strat
egy parameter gives us a number of options for generating values. There are two par‐
ticularly useful strategies. First, stratified makes predictions that are proportional
to the training set’s target vector’s class proportions (i.e., if 20% of the observations in
the training data are women, then DummyClassifier will predict women 20% of the
11.3 Creating a Baseline Classification Model
|
185
time). Second, uniform will generate predictions uniformly at random between the
different classes. For example, if 20% of observations are women and 80% are men,
uniform will produce predictions that are 50% women and 50% men.
See Also
• scikit-learn documentation: DummyClassifier
11.4 Evaluating Binary Classifier Predictions
Problem
Given a trained classification model, you want to evaluate its quality.
Solution
Use scikit-learn’s cross_val_score to conduct cross-validation while using the scor
ing parameter to define one of a number of performance metrics, including accuracy,
precision, recall, and F1.
Accuracy is a common performance metric. It is simply the proportion of observa‐
tions predicted correctly:
Accuracy =
TP + TN
TP + TN + FP + FN
where:
• TP is the number of true positives. Observations that are part of the positive class
(has the disease, purchased the product, etc.) and that we predicted correctly.
• TN is the number of true negatives. Observations that are part of the negative
class (does not have the disease, did not purchase the product, etc.) and that we
predicted correctly.
• FP is the number of false positives. Also called a Type I error. Observations pre‐
dicted to be part of the positive class that are actually part of the negative class.
• FN is the number of false negatives. Also called a Type II error. Observations pre‐
dicted to be part of the negative class that are actually part of the positive class.
We can measure accuracy in three-fold (the default number of folds) cross-validation
by setting scoring="accuracy":
186
|
Chapter 11: Model Evaluation
# Load libraries
from sklearn.model_selection import cross_val_score
from sklearn.linear_model import LogisticRegression
from sklearn.datasets import make_classification
# Generate features matrix and target vector
X, y = make_classification(n_samples = 10000,
n_features = 3,
n_informative = 3,
n_redundant = 0,
n_classes = 2,
random_state = 1)
# Create logistic regression
logit = LogisticRegression()
# Cross-validate model using accuracy
cross_val_score(logit, X, y, scoring="accuracy")
array([ 0.95170966,
0.9580084 ,
0.95558223])
The appeal of accuracy is that it has an intuitive and plain English explanation: pro‐
portion of observations predicted correctly. However, in the real world, often our data
has imbalanced classes (e.g., the 99.9% of observations are of class 1 and only 0.1%
are class 2). When in the presence of imbalanced classes, accuracy suffers from a par‐
adox where a model is highly accurate but lacks predictive power. For example, imag‐
ine we are trying to predict the presence of a very rare cancer that occurs in 0.1% of
the population. After training our model, we find the accuracy is at 95%. However,
99.9% of people do not have the cancer: if we simply created a model that “predicted”
that nobody had that form of cancer, our naive model would be 4.9% more accurate,
but clearly is not able to predict anything. For this reason, we are often motivated to
use other metrics like precision, recall, and the F1 score.
Precision is the proportion of every observation predicted to be positive that is
actually positive. We can think about it as a measurement noise in our predictions—
that is, when we predict something is positive, how likely we are to be right. Models
with high precision are pessimistic in that they only predict an observation is of the
positive class when they are very certain about it. Formally, precision is:
Precision =
TP
TP + FP
# Cross-validate model using precision
cross_val_score(logit, X, y, scoring="precision")
array([ 0.95252404,
0.96583282,
0.95558223])
Recall is the proportion of every positive observation that is truly positive. Recall
measures the model’s ability to identify an observation of the positive class. Models
11.4 Evaluating Binary Classifier Predictions
|
187
with high recall are optimistic in that they have a low bar for predicting that an obser‐
vation is in the positive class:
Recall =
TP
TP + FN
# Cross-validate model using recall
cross_val_score(logit, X, y, scoring="recall")
array([ 0.95080984,
0.94961008,
0.95558223])
If this is the first time you have encountered precision and recall, it is understandable
if it takes you a little while to fully understand them. This is one of the downsides to
accuracy; precision and recall are less intuitive. Almost always we want some kind of
balance between precision and recall, and this role is filled by the F1 score. The F1
score is the harmonic mean (a kind of average used for ratios):
F1 = 2 ×
Precision × Recall
Precision + Recall
It is a measure of correctness achieved in positive prediction—that is, of observations
labeled as positive, how many are actually positive:
# Cross-validate model using f1
cross_val_score(logit, X, y, scoring="f1")
array([ 0.95166617,
0.95765275,
0.95558223])
Discussion
As an evaluation metric, accuracy has some valuable properties, especially its simple
intuition. However, better metrics often involve using some balance of precision and
recall—that is, a trade-off between the optimism and pessimism of our model. F1 rep‐
resents a balance between the recall and precision, where the relative contributions of
both are equal.
Alternatively to using cross_val_score, if we already have the true y values and the
predicted y values, we can calculate metrics like accuracy and recall directly:
# Load library
from sklearn.model_selection import train_test_split
from sklearn.metrics import accuracy_score
# Create training and test split
X_train, X_test, y_train, y_test = train_test_split(X,
y,
test_size=0.1,
random_state=1)
188
|
Chapter 11: Model Evaluation
# Predict values for training target vector
y_hat = logit.fit(X_train, y_train).predict(X_test)
# Calculate accuracy
accuracy_score(y_test, y_hat)
0.94699999999999995
See Also
• Accuracy paradox
11.5 Evaluating Binary Classifier Thresholds
Problem
You want to evaluate a binary classifier and various probability thresholds.
Solution
The Receiving Operating Characteristic (ROC) curve is a common method for evalu‐
ating the quality of a binary classifier. ROC compares the presence of true positives
and false positives at every probability threshold (i.e., the probability at which an
observation is predicted to be a class). By plotting the ROC curve, we can see how the
model performs. A classifier that predicts every observation correctly would look like
the solid light gray line in the following chart, going straight up to the top immedi‐
ately. A classifier that predicts at random will appear as the diagonal line. The better
the model, the closer it is to the solid line. In scikit-learn, we can use roc_curve to
calculate the true and false positives at each threshold, then plot them:
# Load libraries
import matplotlib.pyplot as plt
from sklearn.datasets import make_classification
from sklearn.linear_model import LogisticRegression
from sklearn.metrics import roc_curve, roc_auc_score
from sklearn.model_selection import train_test_split
# Create feature matrix and target vector
features, target = make_classification(n_samples=10000,
n_features=10,
n_classes=2,
n_informative=3,
random_state=3)
# Split into training and test sets
features_train, features_test, target_train, target_test = train_test_split(
features, target, test_size=0.1, random_state=1)
11.5 Evaluating Binary Classifier Thresholds
|
189
# Create classifier
logit = LogisticRegression()
# Train model
logit.fit(features_train, target_train)
# Get predicted probabilities
target_probabilities = logit.predict_proba(features_test)[:,1]
# Create true and false positive rates
false_positive_rate, true_positive_rate, threshold = roc_curve(target_test,
target_probabilities)
# Plot ROC curve
plt.title("Receiver Operating Characteristic")
plt.plot(false_positive_rate, true_positive_rate)
plt.plot([0, 1], ls="--")
plt.plot([0, 0], [1, 0] , c=".7"), plt.plot([1, 1] , c=".7")
plt.ylabel("True Positive Rate")
plt.xlabel("False Positive Rate")
plt.show()
190
|
Chapter 11: Model Evaluation
Discussion
Up until now we have only examined models based on the values they predict. How‐
ever, in many learning algorithms those predicted values are based off of probability
estimates. That is, each observation is given an explicit probability of belonging in
each class. In our solution, we can use predict_proba to see the predicted probabili‐
ties for the first observation:
# Get predicted probabilities
logit.predict_proba(features_test)[0:1]
array([[ 0.8688938,
0.1311062]])
We can see the classes using classes_:
logit.classes_
array([0, 1])
In this example, the first observation has an ~87% chance of being in the negative
class (0) and a 13% chance of being in the positive class (1). By default, scikit-learn
predicts an observation is part of the positive class if the probability is greater than
0.5 (called the threshold). However, instead of a middle ground, we will often want to
explicitly bias our model to use a different threshold for substantive reasons. For
example, if a false positive is very costly to our company, we might prefer a model
that has a high probability threshold. We fail to predict some positives, but when an
observation is predicted to be positive, we can be very confident that the prediction is
correct. This trade-off is represented in the true positive rate (TPR) and the false pos‐
itive rate (FPR). The true positive rate is the number of observations correctly predic‐
ted true divided by all true positive observations:
TPR =
True Positives
True Positives+False Negatives
The false positive rate is the number of incorrectly predicted positives divided by all
true negative observations:
FPR =
False Positives
False Positives+True Negatives
The ROC curve represents the respective TPR and FPR for every probability thres‐
hold. For example, in our solution a threshold of roughly 0.50 has a TPR of \0.81 and an
FPR of \0.15:
print("Threshold:", threshold[116])
print("True Positive Rate:", true_positive_rate[116])
print("False Positive Rate:", false_positive_rate[116])
11.5 Evaluating Binary Classifier Thresholds
|
191
Threshold: 0.528224777887
True Positive Rate: 0.810204081633
False Positive Rate: 0.154901960784
However, if we increase the threshold to ~80% (i.e., increase how certain the model
has to be before it predicts an observation as positive) the TPR drops significantly but
so does the FPR:
print("Threshold:", threshold[45])
print("True Positive Rate:", true_positive_rate[45])
print("False Positive Rate:", false_positive_rate[45])
Threshold: 0.808019566563
True Positive Rate: 0.563265306122
False Positive Rate: 0.0470588235294
This is because our higher requirement for being predicted to be in the positive class
has made the model not identify a number of positive observations (the lower TPR),
but also reduce the noise from negative observations being predicted as positive (the
lower FPR).
In addition to being able to visualize the trade-off between TPR and FPR, the ROC
curve can also be used as a general metric for a model. The better a model is, the
higher the curve and thus the greater the area under the curve. For this reason, it is
common to calculate the area under the ROC curve (AUCROC) to judge the overall
equality of a model at all possible thresholds. The closer the AUCROC is to 1, the bet‐
ter the model. In scikit-learn we can calculate the AUCROC using roc_auc_score:
# Calculate area under curve
roc_auc_score(target_test, target_probabilities)
0.90733893557422962
See Also
• ROC Curves in Python and R
• The Area Under an ROC Curve
11.6 Evaluating Multiclass Classifier Predictions
Problem
You have a model that predicts three or more classes and want to evaluate its perfor‐
mance.
192
|
Chapter 11: Model Evaluation
Solution
Use cross-validation with an evaluation metric capable of handling more than two
classes:
# Load libraries
from sklearn.model_selection import cross_val_score
from sklearn.linear_model import LogisticRegression
from sklearn.datasets import make_classification
# Generate features matrix and target vector
features, target = make_classification(n_samples = 10000,
n_features = 3,
n_informative = 3,
n_redundant = 0,
n_classes = 3,
random_state = 1)
# Create logistic regression
logit = LogisticRegression()
# Cross-validate model using accuracy
cross_val_score(logit, features, target, scoring='accuracy')
array([ 0.83653269,
0.8259826 ,
0.81308131])
Discussion
When we have balanced classes (e.g., a roughly equal number of observations in each
class of the target vector), accuracy is—just like in the binary class setting—a simple
and interpretable choice for an evaluation metric. Accuracy is the number of correct
predictions divided by the number of observations and works just as well in the mul‐
ticlass as binary setting. However, when we have imbalanced classes (a common sce‐
nario), we should be inclined to use other evaluation metrics.
Many of scikit-learn’s built-in metrics are for evaluating binary classifiers. However,
many of these metrics can be extended for use when we have more than two classes.
Precision, recall, and F1 scores are useful metrics that we have already covered in
detail in previous recipes. While all of them were originally designed for binary clas‐
sifiers, we can apply them to multiclass settings by treating our data as a set of binary
classes. Doing so enables us to apply the metrics to each class as if it were the only
class in the data, and then aggregate the evaluation scores for all the classes by averag‐
ing them:
# Cross-validate model using macro averaged F1 score
cross_val_score(logit, features, target, scoring='f1_macro')
array([ 0.83613125,
0.82562258,
0.81293539])
11.6 Evaluating Multiclass Classifier Predictions
|
193
In this code, _macro refers to the method used to average the evaluation scores from
the classes:
macro
Calculate mean of metric scores for each class, weighting each class equally.
weighted
Calculate mean of metric scores for each class, weighting each class proportional
to its size in the data.
micro
Calculate mean of metric scores for each observation-class combination.
11.7 Visualizing a Classifier’s Performance
Problem
Given predicted classes and true classes of the test data, you want to visually compare
the model’s quality.
Solution
Use a confusion matrix, which compares predicted classes and true classes:
# Load libraries
import matplotlib.pyplot as plt
import seaborn as sns
from sklearn import datasets
from sklearn.linear_model import LogisticRegression
from sklearn.model_selection import train_test_split
from sklearn.metrics import confusion_matrix
import pandas as pd
# Load data
iris = datasets.load_iris()
# Create feature matrix
features = iris.data
# Create target vector
target = iris.target
# Create list of target class names
class_names = iris.target_names
# Create training and test set
features_train, features_test, target_train, target_test = train_test_split(
features, target, random_state=1)
194
|
Chapter 11: Model Evaluation
# Create logistic regression
classifier = LogisticRegression()
# Train model and make predictions
target_predicted = classifier.fit(features_train,
target_train).predict(features_test)
# Create confusion matrix
matrix = confusion_matrix(target_test, target_predicted)
# Create pandas dataframe
dataframe = pd.DataFrame(matrix, index=class_names, columns=class_names)
# Create heatmap
sns.heatmap(dataframe, annot=True, cbar=None, cmap="Blues")
plt.title("Confusion Matrix"), plt.tight_layout()
plt.ylabel("True Class"), plt.xlabel("Predicted Class")
plt.show()
Discussion
Confusion matrices are an easy, effective visualization of a classifier’s performance.
One of the major benefits of confusion matrices is their interpretability. Each column
of the matrix (often visualized as a heatmap) represents predicted classes, while every
row shows true classes. The end result is that every cell is one possible combination of
predict and true classes. This is probably best explained using an example. In the
11.7 Visualizing a Classifier’s Performance
|
195
solution, the top-left cell is the number of observations predicted to be Iris setosa
(indicated by the column) that are actually Iris setosa (indicated by the row). This
means the models accurately predicted all Iris setosa flowers. However, the model
does not do as well at predicting Iris virginica. The bottom-right cell indicates that the
model successfully predicted nine observations were Iris virginica, but (looking one
cell up) predicted six flowers to be viriginica that were actually Iris versicolor.
There are three things worth noting about confusion matrices. First, a perfect model
will have values along the diagonal and zeros everywhere else. A bad model will look
like the observation counts will be spread evenly around cells. Second, a confusion
matrix lets us see not only where the model was wrong, but also how it was wrong.
That is, we can look at patterns of misclassification. For example, our model had an
easy time differentiating Iris virginica and Iris setosa, but a much more difficult time
classifying Iris virginica and Iris versicolor. Finally, confusion matrices work with any
number of classes (although if we had one million classes in our target vector, the
confusion matrix visualization might be difficult to read).
See Also
• Confusion Matrix
• scikit-learn documentation: Confusion Matrix
11.8 Evaluating Regression Models
Problem
You want to evaluate the performance of a regression model.
Solution
Use mean squared error (MSE):
# Load libraries
from sklearn.datasets import make_regression
from sklearn.model_selection import cross_val_score
from sklearn.linear_model import LinearRegression
# Generate features matrix, target vector
features, target = make_regression(n_samples = 100,
n_features = 3,
n_informative = 3,
n_targets = 1,
noise = 50,
coef = False,
random_state = 1)
196
|
Chapter 11: Model Evaluation
# Create a linear regression object
ols = LinearRegression()
# Cross-validate the linear regression using (negative) MSE
cross_val_score(ols, features, target, scoring='neg_mean_squared_error')
array([-1718.22817783, -3103.4124284 , -1377.17858823])
Another common regression metric is the coefficient of determination, R2:
# Cross-validate the linear regression using R-squared
cross_val_score(ols, features, target, scoring='r2')
array([ 0.87804558,
0.76395862,
0.89154377])
Discussion
MSE is one of the most common evaluation metrics for regression models. Formally,
MSE is:
MSE =
1 n
yi − yi
ni∑
=1
2
where n is the number of observations, yi is the true value of the target we are trying
to predict for observation i, and yi is the model’s predicted value for yi. MSE is a
measurement of the squared sum of all distances between predicted and true values.
The higher the value of MSE, the greater the total squared error and thus the worse
the model. There are a number of mathematical benefits to squaring the error term,
including that it forces all error values to be positive, but one often unrealized impli‐
cation is that squaring penalizes a few large errors more than many small errors, even
if the absolute value of the errors is the same. For example, imagine two models, A
and B, each with two observations:
• Model A has errors of 0 and 10 and thus its MSE is 02 + 102 = 100.
• Model B has two errors of 5 each, and thus its MSE is 52 + 52 = 50.
Both models have the same total error, 10; however, MSE would consider Model A
(MSE = 100) worse than Model B (MSE = 50). In practice this implication is rarely an
issue (and indeed can be theoretically beneficial) and MSE works perfectly fine as an
evaluation metric.
One important note: by default in scikit-learn arguments of the scoring parameter
assume that higher values are better than lower values. However, this is not the case
for MSE, where higher values mean a worse model. For this reason, scikit-learn looks
at the negative MSE using the neg_mean_squared_error argument.
11.8 Evaluating Regression Models
|
197
A common alternative regression evaluation metric is R2, which measures the amount
of variance in the target vector that is explained by the model:
R2 = 1 −
∑ni = 1 yi − y i
∑ni = 1 yi − y
2
2
where yi is the true target value of the ith observation, y i is the predicted value for the
ith observation, and y is the mean value of the target vector. The closer to 1.0, the
better the model.
See Also
• Mean squared error
• Coefficient of determination
11.9 Evaluating Clustering Models
Problem
You have used an unsupervised learning algorithm to cluster your data. Now you
want to know how well it did.
Solution
The short answer is that you probably can’t, at least not in the way you want.
That said, one option is to evaluate clustering using silhouette coefficients, which
measure the quality of the clusters:
import numpy as np
from sklearn.metrics import silhouette_score
from sklearn import datasets
from sklearn.cluster import KMeans
from sklearn.datasets import make_blobs
# Generate feature matrix
features, _ = make_blobs(n_samples = 1000,
n_features = 10,
centers = 2,
cluster_std = 0.5,
shuffle = True,
random_state = 1)
# Cluster data using k-means to predict classes
198
|
Chapter 11: Model Evaluation
model = KMeans(n_clusters=2, random_state=1).fit(features)
# Get predicted classes
target_predicted = model.labels_
# Evaluate model
silhouette_score(features, target_predicted)
0.89162655640721422
Discussion
Supervised model evaluation compares predictions (e.g., classes or quantitative val‐
ues) with the corresponding true values in the target vector. However, the most com‐
mon motivation for using clustering methods is that your data doesn’t have a target
vector. There are a number of clustering evaluation metrics that require a target vec‐
tor, but again, using unsupervised learning approaches like clustering when you have
a target vector available to you is probably handicapping yourself unnecessarily.
While we cannot evaluate predictions versus true values if we don’t have a target vec‐
tor, we can evaluate the nature of the clusters themselves. Intuitively, we can imagine
“good” clusters having very small distances between observations in the same cluster
(i.e., dense clusters) and large distances between the different clusters (i.e., wellseparated clusters). Silhouette coefficients provide a single value measuring both
traits. Formally, the ith observation’s silhouette coefficient is:
si =
bi − ai
max ai, bi
where si is the silhouette coefficient for observation i, ai is the mean distance between
i and all observations of the same class, and bi is the mean distance between i and all
observations from the closest cluster of a different class. The value returned by sil
houette_score is the mean silhouette coefficient for all observations. Silhouette coef‐
ficients range between –1 and 1, with 1 indicating dense, well-separated clusters.
See Also
• scikit-learn documentation: silhouette_score
11.10 Creating a Custom Evaluation Metric
Problem
You want to evaluate a model using a metric you created.
11.10 Creating a Custom Evaluation Metric
|
199
Solution
Create the metric as a function and convert it into a scorer function using scikitlearn’s make_scorer:
# Load libraries
from sklearn.metrics import make_scorer, r2_score
from sklearn.model_selection import train_test_split
from sklearn.linear_model import Ridge
from sklearn.datasets import make_regression
# Generate features matrix and target vector
features, target = make_regression(n_samples = 100,
n_features = 3,
random_state = 1)
# Create training set and test set
features_train, features_test, target_train, target_test = train_test_split(
features, target, test_size=0.10, random_state=1)
# Create custom metric
def custom_metric(target_test, target_predicted):
# Calculate r-squared score
r2 = r2_score(target_test, target_predicted)
# Return r-squared score
return r2
# Make scorer and define that higher scores are better
score = make_scorer(custom_metric, greater_is_better=True)
# Create ridge regression object
classifier = Ridge()
# Train ridge regression model
model = classifier.fit(features_train, target_train)
# Apply custom scorer
score(model, features_test, target_test)
0.99979061028820582
Discussion
While scikit-learn has a number of built-in metrics for evaluating model perfor‐
mance, it is often useful to define our own metrics. scikit-learn makes this easy using
make_scorer. First, we define a function that takes in two arguments—the ground
truth target vector and our predicted values—and outputs some score. Second, we
use make_scorer to create a scorer object, making sure to specify whether higher or
lower scores are desirable (using the greater_is_better parameter).
200
|
Chapter 11: Model Evaluation
The custom metric in the solution (custom_metric) is a toy example since it simply
wraps a built-in metric for calculating the R2 score. In a real-world situation, we
would replace the custom_metric function with whatever custom metric we wanted.
However, we can see that the custom metric that calculates R2 does work by compar‐
ing the results to scikit-learn’s r2_score built-in method:
# Predict values
target_predicted = model.predict(features_test)
# Calculate r-squared score
r2_score(target_test, target_predicted)
0.99979061028820582
See Also
• scikit-learn documentation: make_scorer
11.11 Visualizing the Effect of Training Set Size
Problem
You want to evaluate the effect of the number of observations in your training set on
some metric (accuracy, F1, etc.).
Solution
Plot the learning curve:
# Load libraries
import numpy as np
import matplotlib.pyplot as plt
from sklearn.ensemble import RandomForestClassifier
from sklearn.datasets import load_digits
from sklearn.model_selection import learning_curve
# Load data
digits = load_digits()
# Create feature matrix and target vector
features, target = digits.data, digits.target
# Create CV training and test scores for various training set sizes
train_sizes, train_scores, test_scores = learning_curve(# Classifier
RandomForestClassifier(),
# Feature matrix
features,
# Target vector
11.11 Visualizing the Effect of Training Set Size
|
201
target,
# Number of folds
cv=10,
# Performance metric
scoring='accuracy',
# Use all computer cores
n_jobs=-1,
# Sizes of 50
# training set
train_sizes=np.linspace(
0.01,
1.0,
50))
# Create means and standard deviations of training set scores
train_mean = np.mean(train_scores, axis=1)
train_std = np.std(train_scores, axis=1)
# Create means and standard deviations of test set scores
test_mean = np.mean(test_scores, axis=1)
test_std = np.std(test_scores, axis=1)
# Draw lines
plt.plot(train_sizes, train_mean, '--', color="#111111", label="Training score")
plt.plot(train_sizes, test_mean, color="#111111", label="Cross-validation score")
# Draw bands
plt.fill_between(train_sizes, train_mean - train_std,
train_mean + train_std, color="#DDDDDD")
plt.fill_between(train_sizes, test_mean - test_std,
test_mean + test_std, color="#DDDDDD")
# Create plot
plt.title("Learning Curve")
plt.xlabel("Training Set Size"), plt.ylabel("Accuracy Score"),
plt.legend(loc="best")
plt.tight_layout()
plt.show()
202
|
Chapter 11: Model Evaluation
Discussion
Learning curves visualize the performance (e.g., accuracy, recall) of a model on the
training set and during cross-validation as the number of observations in the training
set increases. They are commonly used to determine if our learning algorithms would
benefit from gathering additional training data.
In our solution, we plot the accuracy of a random forest classifier at 50 different
training set sizes ranging from 1% of observations to 100%. The increasing accuracy
score of the cross-validated models tell us that we would likely benefit from addi‐
tional observations (although in practice this might not be feasible).
See Also
• scikit-learn documentation: Learning Curve
11.12 Creating a Text Report of Evaluation Metrics
Problem
You want a quick description of a classifier’s performance.
11.12 Creating a Text Report of Evaluation Metrics
|
203
Solution
Use scikit-learn’s classification_report:
# Load libraries
from sklearn import datasets
from sklearn.linear_model import LogisticRegression
from sklearn.model_selection import train_test_split
from sklearn.metrics import classification_report
# Load data
iris = datasets.load_iris()
# Create feature matrix
features = iris.data
# Create target vector
target = iris.target
# Create list of target class names
class_names = iris.target_names
# Create training and test set
features_train, features_test, target_train, target_test = train_test_split(
features, target, random_state=1)
# Create logistic regression
classifier = LogisticRegression()
# Train model and make predictions
model = classifier.fit(features_train, target_train)
target_predicted = model.predict(features_test)
# Create a classification report
print(classification_report(target_test,
target_predicted,
target_names=class_names))
precision
recall
f1-score
support
setosa
versicolor
virginica
1.00
1.00
0.60
1.00
0.62
1.00
1.00
0.77
0.75
13
16
9
avg / total
0.91
0.84
0.84
38
Discussion
classification_report provides a quick means for us to see some common evalua‐
tion metrics, including precision, recall, and F1-score (described earlier in this chap‐
ter). Support refers to the number of observations in each class.
204
|
Chapter 11: Model Evaluation
See Also
• Precision and recall
11.13 Visualizing the Effect of Hyperparameter Values
Problem
You want to understand how the performance of a model changes as the value of
some hyperparameter changes.
Solution
Plot the validation curve:
# Load libraries
import matplotlib.pyplot as plt
import numpy as np
from sklearn.datasets import load_digits
from sklearn.ensemble import RandomForestClassifier
from sklearn.model_selection import validation_curve
# Load data
digits = load_digits()
# Create feature matrix and target vector
features, target = digits.data, digits.target
# Create range of values for parameter
param_range = np.arange(1, 250, 2)
# Calculate accuracy on training and test set using range of parameter values
train_scores, test_scores = validation_curve(
# Classifier
RandomForestClassifier(),
# Feature matrix
features,
# Target vector
target,
# Hyperparameter to examine
param_name="n_estimators",
# Range of hyperparameter's values
param_range=param_range,
# Number of folds
cv=3,
# Performance metric
scoring="accuracy",
# Use all computer cores
n_jobs=-1)
11.13 Visualizing the Effect of Hyperparameter Values
|
205
# Calculate mean and standard deviation for training set scores
train_mean = np.mean(train_scores, axis=1)
train_std = np.std(train_scores, axis=1)
# Calculate mean and standard deviation for test set scores
test_mean = np.mean(test_scores, axis=1)
test_std = np.std(test_scores, axis=1)
# Plot mean accuracy scores for training and test sets
plt.plot(param_range, train_mean, label="Training score", color="black")
plt.plot(param_range, test_mean, label="Cross-validation score", color="dimgrey")
# Plot accurancy bands for training and test sets
plt.fill_between(param_range, train_mean - train_std,
train_mean + train_std, color="gray")
plt.fill_between(param_range, test_mean - test_std,
test_mean + test_std, color="gainsboro")
# Create plot
plt.title("Validation Curve With Random Forest")
plt.xlabel("Number Of Trees")
plt.ylabel("Accuracy Score")
plt.tight_layout()
plt.legend(loc="best")
plt.show()
206
|
Chapter 11: Model Evaluation
Discussion
Most training algorithms (including many covered in this book) contain hyperpara‐
meters that must be chosen before the training process begins. For example, a ran‐
dom forest classifier creates a “forest” of decision trees, each of which votes on the
predicted class of an observation. One hyperparameter in random forest classifiers is
the number of trees in the forest. Most often hyperparameter values are selected dur‐
ing model selection (see Chapter 12). However, it is occasionally useful to visualize
how model performance changes as the hyperparameter value changes. In our solu‐
tion, we plot the changes in accuracy for a random forest classifier for the training set
and during cross-validation as the number of trees increases. When we have a small
number of trees, both the training and cross-validation score are low, suggesting the
model is underfitted. As the number of trees increases to 250, the accuracy of both
levels off, suggesting there is probably not much value in the computational cost of
training a massive forest.
In scikit-learn, we can calculate the validation curve using validation_curve, which
contains three important parameters:
• param_name is the name of the hyperparameter to vary.
• param_range is the value of the hyperparameter to use.
• scoring is the evaluation metric used to judge to model.
See Also
• scikit-learn documentation: Validation Curve
11.13 Visualizing the Effect of Hyperparameter Values
|
207
CHAPTER 12
Model Selection
12.0 Introduction
In machine learning, we use training algorithms to learn the parameters of a model
by minimizing some loss function. However, in addition many learning algorithms
(e.g., support vector classifier and random forests) also have hyperparameters that
must be defined outside of the learning process. For example, random forests are col‐
lections of decision trees (hence the word forest); however, the number of decision
trees in the forest is not learned by the algorithm and must be set prior to fitting. This
is often referred to as hyperparameter tuning, hyperparameter optimization, or model
selection. Additionally, often we might want to try multiple learning algorithms (for
example, trying both support vector classifier and random forests to see which learn‐
ing method produces the best model).
While there is widespread variation in the use of terminology in this area, in this
book we refer to both selecting the best learning algorithm and its best hyperparame‐
ters as model selection. The reason is straightforward: imagine we have data and want
to train a support vector classifier with 10 candidate hyperparameter values and a
random forest classifier with 10 candidate hyperparameter values. The result is that
we are trying to select the best model from a set of 20 candidate models. In this chap‐
ter, we will cover techniques to efficiently select the best model from the set of candi‐
dates.
Throughout this chapter we will refer to specific hyperparameters, such as C (the
inverse of regularization strength). Do not worry if you don’t know what the hyper‐
parameters are. We will cover them in later chapters. Instead, just treat hyperparame‐
ters like the settings for the learning algorithm that we must choose before starting
training.
209
12.1 Selecting Best Models Using Exhaustive Search
Problem
You want to select the best model by searching over a range of hyperparameters.
Solution
Use scikit-learn’s GridSearchCV:
# Load libraries
import numpy as np
from sklearn import linear_model, datasets
from sklearn.model_selection import GridSearchCV
# Load data
iris = datasets.load_iris()
features = iris.data
target = iris.target
# Create logistic regression
logistic = linear_model.LogisticRegression()
# Create range of candidate penalty hyperparameter values
penalty = ['l1', 'l2']
# Create range of candidate regularization hyperparameter values
C = np.logspace(0, 4, 10)
# Create dictionary hyperparameter candidates
hyperparameters = dict(C=C, penalty=penalty)
# Create grid search
gridsearch = GridSearchCV(logistic, hyperparameters, cv=5, verbose=0)
# Fit grid search
best_model = gridsearch.fit(features, target)
Discussion
GridSearchCV is a brute-force approach to model selection using cross-validation.
Specifically, a user defines sets of possible values for one or multiple hyperparameters,
and then GridSearchCV trains a model using every value and/or combination of val‐
ues. The model with the best performance score is selected as the best model. For
example, in our solution we used logistic regression as our learning algorithm, con‐
taining two hyperparameters: C and the regularization penalty. Don’t worry if you
don’t know what C and regularization mean; we cover them in the next few chapters.
Just realize that C and the regularization penalty can take a range of values, which
have to be specified prior to training. For C, we define 10 possible values:
210
|
Chapter 12: Model Selection
np.logspace(0, 4, 10)
array([
1.00000000e+00,
2.15443469e+01,
4.64158883e+02,
1.00000000e+04])
2.78255940e+00,
5.99484250e+01,
1.29154967e+03,
7.74263683e+00,
1.66810054e+02,
3.59381366e+03,
And similarly we define two possible values for the regularization penalty: ['l1',
'l2']. For each combination of C and regularization penalty values, we train the
model and evaluate it using k-fold cross-validation. In our solution, we had 10 possi‐
ble values of C, 2 possible values of regularization penalty, and 5 folds. They created
10 × 2 × 5 = 100 candidate models from which the best was selected.
Once GridSearchCV is complete, we can see the hyperparameters of the best model:
# View best hyperparameters
print('Best Penalty:', best_model.best_estimator_.get_params()['penalty'])
print('Best C:', best_model.best_estimator_.get_params()['C'])
Best Penalty: l1
Best C: 7.74263682681
By default, after identifying the best hyperparameters, GridSearchCV will retrain a
model using the best hyperparameters on the entire dataset (rather than leaving a fold
out for cross-validation). We can use this model to predict values like any other
scikit-learn model:
# Predict target vector
best_model.predict(features)
array([0,
0,
0,
1,
1,
2,
2,
0,
0,
0,
2,
1,
2,
2,
0,
0,
0,
1,
1,
2,
2,
0,
0,
0,
1,
1,
2,
2,
0,
0,
1,
1,
1,
2,
2,
0,
0,
1,
1,
1,
2,
2,
0,
0,
1,
1,
1,
2,
2,
0,
0,
1,
1,
1,
2,
2,
0,
0,
1,
1,
2,
2,
2,
0,
0,
1,
1,
2,
2,
2,
0,
0,
1,
1,
2,
2,
2,
0, 0,
0, 0,
1, 1,
1, 1,
2, 2,
2, 2,
2])
0,
0,
1,
1,
2,
2,
0,
0,
1,
2,
2,
2,
0,
0,
1,
1,
2,
2,
0,
0,
1,
1,
2,
2,
0,
0,
1,
1,
2,
2,
0,
0,
1,
1,
2,
1,
0,
0,
1,
1,
2,
2,
0,
0,
1,
1,
2,
2,
0,
0,
1,
1,
2,
2,
0,
0,
1,
1,
2,
2,
One GridSearchCV parameter is worth noting: verbose. While mostly unnecessary, it
can be reassuring during long searching processes to receive an indication that the
search is progressing. The verbose parameter determines the amount of messages
outputted during the search, with 0 showing no output, and 1 to 3 outputting mes‐
sages with increasing detail.
See Also
• scikit-learn documentation: GridSearchCV
12.1 Selecting Best Models Using Exhaustive Search
|
211
12.2 Selecting Best Models Using Randomized Search
Problem
You want a computationally cheaper method than exhaustive search to select the best
model.
Solution
Use scikit-learn’s RandomizedSearchCV:
# Load libraries
from scipy.stats import uniform
from sklearn import linear_model, datasets
from sklearn.model_selection import RandomizedSearchCV
# Load data
iris = datasets.load_iris()
features = iris.data
target = iris.target
# Create logistic regression
logistic = linear_model.LogisticRegression()
# Create range of candidate regularization penalty hyperparameter values
penalty = ['l1', 'l2']
# Create distribution of candidate regularization hyperparameter values
C = uniform(loc=0, scale=4)
# Create hyperparameter options
hyperparameters = dict(C=C, penalty=penalty)
# Create randomized search
randomizedsearch = RandomizedSearchCV(
logistic, hyperparameters, random_state=1, n_iter=100, cv=5, verbose=0,
n_jobs=-1)
# Fit randomized search
best_model = randomizedsearch.fit(features, target)
Discussion
In Recipe 12.1, we used GridSearchCV on a user-defined set of hyperparameter values
to search for the best model according to a score function. A more efficient method
than GridSearchCV’s brute-force search is to search over a specific number of random
combinations of hyperparameter values from user-supplied distributions (e.g., nor‐
mal, uniform). scikit-learn implements this randomized search technique with Ran
domizedSearchCV.
212
|
Chapter 12: Model Selection
With RandomizedSearchCV, if we specify a distribution, scikit-learn will randomly
sample without replacement hyperparameter values from that distribution. As an
example of the general concept, here we randomly sample 10 values from a uniform
distribution ranging from 0 to 4:
# Define a uniform distribution between 0 and 4, sample 10 values
uniform(loc=0, scale=4).rvs(10)
array([ 1.97118721,
1.20394179,
1.34786898,
0.88117832,
3.79867719,
0.47855329,
1.15999967,
0.48521699,
3.33485021,
1.5479854 ])
Alternatively, if we specify a list of values such as two regularization penalty hyper‐
parameter values, ['l1', 'l2'], RandomizedSearchCV will randomly sample with
replacement from the list.
Just like with GridSearchCV, we can see the hyperparameter values of the best model:
# View best hyperparameters
print('Best Penalty:', best_model.best_estimator_.get_params()['penalty'])
print('Best C:', best_model.best_estimator_.get_params()['C'])
Best Penalty: l1
Best C: 1.66808801881
And just like with GridSearchCV, after the search is complete RandomizedSearchCV
fits a new model using the best hyperparameters on the entire dataset. We can use this
model like any other in scikit-learn; for example, to make predictions:
# Predict target vector
best_model.predict(features)
array([0,
0,
0,
1,
1,
2,
2,
0,
0,
0,
2,
1,
2,
2,
0,
0,
0,
1,
1,
2,
2,
0,
0,
0,
1,
1,
2,
2,
0,
0,
1,
1,
1,
2,
2,
0,
0,
1,
1,
1,
2,
2,
0,
0,
1,
1,
1,
2,
2,
0,
0,
1,
1,
1,
2,
2,
0,
0,
1,
1,
2,
2,
2,
0,
0,
1,
1,
2,
2,
2,
0,
0,
1,
1,
2,
2,
2,
0, 0,
0, 0,
1, 1,
1, 1,
2, 2,
2, 2,
2])
0,
0,
1,
1,
2,
2,
0,
0,
1,
2,
2,
1,
0,
0,
1,
2,
2,
2,
0,
0,
1,
1,
2,
2,
0,
0,
1,
1,
2,
2,
0,
0,
1,
1,
2,
1,
0,
0,
1,
1,
2,
2,
0,
0,
1,
1,
2,
2,
0,
0,
1,
1,
2,
2,
0,
0,
1,
1,
2,
2,
The number of sampled combinations of hyperparameters (i.e., the number of candi‐
date models trained) is specified with the n_iter (number of iterations) setting.
See Also
• scikit-learn documentation: RandomizedSearchCV
• Random Search for Hyper-Parameter Optimization
12.2 Selecting Best Models Using Randomized Search
|
213
12.3 Selecting Best Models from Multiple Learning
Algorithms
Problem
You want to select the best model by searching over a range of learning algorithms
and their respective hyperparameters.
Solution
Create a dictionary of candidate learning algorithms and their hyperparameters:
# Load libraries
import numpy as np
from sklearn import datasets
from sklearn.linear_model import LogisticRegression
from sklearn.ensemble import RandomForestClassifier
from sklearn.model_selection import GridSearchCV
from sklearn.pipeline import Pipeline
# Set random seed
np.random.seed(0)
# Load data
iris = datasets.load_iris()
features = iris.data
target = iris.target
# Create a pipeline
pipe = Pipeline([("classifier", RandomForestClassifier())])
# Create dictionary with candidate learning algorithms and their hyperparameters
search_space = [{"classifier": [LogisticRegression()],
"classifier__penalty": ['l1', 'l2'],
"classifier__C": np.logspace(0, 4, 10)},
{"classifier": [RandomForestClassifier()],
"classifier__n_estimators": [10, 100, 1000],
"classifier__max_features": [1, 2, 3]}]
# Create grid search
gridsearch = GridSearchCV(pipe, search_space, cv=5, verbose=0)
# Fit grid search
best_model = gridsearch.fit(features, target)
Discussion
In the previous two recipes we found the best model by searching over possible
hyperparameter values of a learning algorithm. However, what if we are not certain
214
|
Chapter 12: Model Selection
which learning algorithm to use? Recent versions of scikit-learn allow us to include
learning algorithms as part of the search space. In our solution we define a search
space that includes two learning algorithms: logistic regression and random forest
classifier. Each learning algorithm has its own hyperparameters, and we define their
candidate values using the format classifier__[hyperparameter name]. For exam‐
ple, for our logistic regression, to define the set of possible values for regularization
hyperparameter space, C, and potential types of regularization penalties, penalty, we
create a dictionary:
{'classifier': [LogisticRegression()],
'classifier__penalty': ['l1', 'l2'],
'classifier__C': np.logspace(0, 4, 10)}
We can also create a similar dictionary for the random forest hyperparameters:
{'classifier': [RandomForestClassifier()],
'classifier__n_estimators': [10, 100, 1000],
'classifier__max_features': [1, 2, 3]}
After the search is complete, we can use best_estimator_ to view the best model’s
learning algorithm and hyperparameters:
# View best model
best_model.best_estimator_.get_params()["classifier"]
LogisticRegression(C=7.7426368268112693, class_weight=None, dual=False,
fit_intercept=True, intercept_scaling=1, max_iter=100,
multi_class='ovr', n_jobs=1, penalty='l1', random_state=None,
solver='liblinear', tol=0.0001, verbose=0, warm_start=False)
Just like with the last two recipes, once we have fit the model selection search, we can
use this best model just like any other scikit-learn model:
# Predict target vector
best_model.predict(features)
array([0,
0,
0,
1,
1,
2,
2,
0,
0,
0,
2,
1,
2,
2,
0,
0,
0,
1,
1,
2,
2,
0,
0,
0,
1,
1,
2,
2,
0,
0,
1,
1,
1,
2,
2,
0,
0,
1,
1,
1,
2,
2,
0,
0,
1,
1,
1,
2,
2,
0,
0,
1,
1,
1,
2,
2,
0,
0,
1,
1,
2,
2,
2,
0,
0,
1,
1,
2,
2,
2,
0,
0,
1,
1,
2,
2,
2,
0, 0,
0, 0,
1, 1,
1, 1,
2, 2,
2, 2,
2])
0,
0,
1,
1,
2,
2,
0,
0,
1,
2,
2,
2,
0,
0,
1,
1,
2,
2,
0,
0,
1,
1,
2,
2,
0,
0,
1,
1,
2,
2,
0,
0,
1,
1,
2,
1,
0,
0,
1,
1,
2,
2,
0,
0,
1,
1,
2,
2,
0,
0,
1,
1,
2,
2,
0,
0,
1,
1,
2,
2,
12.4 Selecting Best Models When Preprocessing
Problem
You want to include a preprocessing step during model selection.
12.4 Selecting Best Models When Preprocessing
|
215
Solution
Create a pipeline that includes the preprocessing step and any of its parameters:
# Load libraries
import numpy as np
from sklearn import datasets
from sklearn.linear_model import LogisticRegression
from sklearn.model_selection import GridSearchCV
from sklearn.pipeline import Pipeline, FeatureUnion
from sklearn.decomposition import PCA
from sklearn.preprocessing import StandardScaler
# Set random seed
np.random.seed(0)
# Load data
iris = datasets.load_iris()
features = iris.data
target = iris.target
# Create a preprocessing object that includes StandardScaler features and PCA
preprocess = FeatureUnion([("std", StandardScaler()), ("pca", PCA())])
# Create a pipeline
pipe = Pipeline([("preprocess", preprocess),
("classifier", LogisticRegression())])
# Create space of candidate values
search_space = [{"preprocess__pca__n_components": [1, 2, 3],
"classifier__penalty": ["l1", "l2"],
"classifier__C": np.logspace(0, 4, 10)}]
# Create grid search
clf = GridSearchCV(pipe, search_space, cv=5, verbose=0, n_jobs=-1)
# Fit grid search
best_model = clf.fit(features, target)
Discussion
Very often we will need to preprocess our data before using it to train a model. We
have to be careful to properly handle preprocessing when conducting model selec‐
tion. First, GridSearchCV uses cross-validation to determine which model has the
highest performance. However, in cross-validation we are in effect pretending that
the fold held out, as the test set is not seen, and thus not part of fitting any prepro‐
cessing steps (e.g., scaling or standardization). For this reason, we cannot preprocess
the data and then run GridSearchCV. Rather, the preprocessing steps must be a part
of the set of actions taken by GridSearchCV. While this might appear complex, the
reality is that scikit-learn makes it simple. FeatureUnion allows us to combine multi‐
216
|
Chapter 12: Model Selection
ple preprocessing actions properly. In our solution we use FeatureUnion to combine
two preprocessing steps: standardize the feature values (StandardScaler) and Princi‐
pal Component Analysis (PCA). This object is called preprocess and contains both of
our preprocessing steps. We then include preprocess into a pipeline with our learn‐
ing algorithm. The end result is that this allows us to outsource the proper (and con‐
fusing) handling of fitting, transforming, and training the models with combinations
of hyperparameters to scikit-learn.
Second, some preprocessing methods have their own parameters, which often have to
be supplied by the user. For example, dimensionality reduction using PCA requires
the user to define the number of principal components to use to produce the trans‐
formed feature set. Ideally, we would choose the number of components that pro‐
duces a model with the greatest performance for some evaluation test metric. Luckily,
scikit-learn makes this easy. When we include candidate component values in the
search space, they are treated like any other hyperparameter to be searched over. In
our solution, we defined features__pca__n_components': [1, 2, 3] in the search
space to indicate that we wanted to discover if one, two, or three principal compo‐
nents produced the best model.
After model selection is complete, we can view the preprocessing values that pro‐
duced the best model. For example, we can see the best number of principal compo‐
nents:
# View best model
best_model.best_estimator_.get_params()['preprocess__pca__n_components']
1
12.5 Speeding Up Model Selection with Parallelization
Problem
You need to speed up model selection.
Solution
Use all the cores in your machine by setting n_jobs=-1:
# Load libraries
import numpy as np
from sklearn import linear_model, datasets
from sklearn.model_selection import GridSearchCV
# Load data
iris = datasets.load_iris()
features = iris.data
target = iris.target
12.5 Speeding Up Model Selection with Parallelization
|
217
# Create logistic regression
logistic = linear_model.LogisticRegression()
# Create range of candidate regularization penalty hyperparameter values
penalty = ["l1", "l2"]
# Create range of candidate values for C
C = np.logspace(0, 4, 1000)
# Create hyperparameter options
hyperparameters = dict(C=C, penalty=penalty)
# Create grid search
gridsearch = GridSearchCV(logistic, hyperparameters, cv=5, n_jobs=-1, verbose=1)
# Fit grid search
best_model = gridsearch.fit(features, target)
Fitting 5 folds for each of 2000 candidates, totalling 10000 fits
[Parallel(n_jobs=-1)]: Done 1352 tasks
| elapsed:
7.2s
[Parallel(n_jobs=-1)]: Done 5120 tasks
| elapsed:
30.3s
[Parallel(n_jobs=-1)]: Done 10000 out of 10000 | elapsed: 1.1min finished
Discussion
In the recipes of this chapter, we have kept the number of candidate models small to
make the code complete quickly. However, in the real world we will often have many
thousands or tens of thousands of models to train. The end result is that it can take
many hours to find the best model. To speed up the process, scikit-learn lets us train
multiple models simultaneously. Without going into too much technical detail, scikitlearn can simultaneously train models up to the number of cores on the machine.
Most modern laptops have four cores, so (assuming you are currently on a laptop) we
can potentially train four models at the same time. This will dramatically increase the
speed of our model selection process. The parameter n_jobs defines the number of
models to train in parallel.
In our solution, we set n_jobs to -1, which tells scikit-learn to use all cores. However,
by default n_jobs is set to 1, meaning it only uses one core. To demonstrate this, if we
run the same GridSearch as in the solution, but with n_jobs=1, we can see it takes
significantly longer to find the best model (note that exact time will depend on your
computer):
# Create grid search using one core
clf = GridSearchCV(logistic, hyperparameters, cv=5, n_jobs=1, verbose=1)
# Fit grid search
best_model = clf.fit(features, target)
218
|
Chapter 12: Model Selection
Fitting 5 folds for each of 2000 candidates, totalling 10000 fits
[Parallel(n_jobs=1)]: Done 10000 out of 10000 | elapsed:
2.8min finished
12.6 Speeding Up Model Selection Using AlgorithmSpecific Methods
Problem
You need to speed up model selection.
Solution
If you are using a select number of learning algorithms, use scikit-learn’s modelspecific cross-validation hyperparameter tuning. For example, LogisticRegres
sionCV:
# Load libraries
from sklearn import linear_model, datasets
# Load data
iris = datasets.load_iris()
features = iris.data
target = iris.target
# Create cross-validated logistic regression
logit = linear_model.LogisticRegressionCV(Cs=100)
# Train model
logit.fit(features, target)
LogisticRegressionCV(Cs=100, class_weight=None, cv=None, dual=False,
fit_intercept=True, intercept_scaling=1.0, max_iter=100,
multi_class='ovr', n_jobs=1, penalty='l2', random_state=None,
refit=True, scoring=None, solver='lbfgs', tol=0.0001, verbose=0)
Discussion
Sometimes the characteristics of a learning algorithm allow us to search for the best
hyperparameters significantly faster than either brute force or randomized model
search methods. In scikit-learn, many learning algorithms (e.g., ridge, lasso, and elas‐
tic net regression) have an algorithm-specific cross-validation method to take advan‐
tage of this. For example, LogisticRegression is used to conduct a standard logistic
regression classifier, while LogisticRegressionCV implements an efficient crossvalidated logistic regression classifier that has the ability to identify the optimum
value of the hyperparameter C.
12.6 Speeding Up Model Selection Using Algorithm-Specific Methods
|
219
scikit-learn’s LogisticRegressionCV method includes a parameter Cs. If supplied a
list, Cs is the candidate hyperparameter values to select from. If supplied an integer,
the parameter Cs generates a list of that many candidate values. The candidate values
are drawn logarithmically from a range between 0.0001 and 1,0000 (a range of rea‐
sonable values for C).
However, a major downside to LogisticRegressionCV is that it can only search a
range of values for C. In Recipe 12.1 our possible hyperparameter space included
both C and another hyperparameter (the regularization penalty norm). This limita‐
tion is common to many of scikit-learn’s model-specific cross-validated approaches.
See Also
• scikit-learn documentation: LogisticRegressionCV
• scikit-learn documentation: Model specific cross-validation
12.7 Evaluating Performance After Model Selection
Problem
You want to evaluate the performance of a model found through model selection.
Solution
Use nested cross-validation to avoid biased evaluation:
# Load libraries
import numpy as np
from sklearn import linear_model, datasets
from sklearn.model_selection import GridSearchCV, cross_val_score
# Load data
iris = datasets.load_iris()
features = iris.data
target = iris.target
# Create logistic regression
logistic = linear_model.LogisticRegression()
# Create range of 20 candidate values for C
C = np.logspace(0, 4, 20)
# Create hyperparameter options
hyperparameters = dict(C=C)
# Create grid search
220
|
Chapter 12: Model Selection
gridsearch = GridSearchCV(logistic, hyperparameters, cv=5, n_jobs=-1, verbose=0)
# Conduct nested cross-validation and outut the average score
cross_val_score(gridsearch, features, target).mean()
0.95343137254901966
Discussion
Nested cross-validation during model selection is a difficult concept for many people
to grasp the first time. Remember that in k-fold cross-validation, we train our model
on k–1 folds of the data, use this model to make predictions on the remaining fold,
and then evaluate our model best on how well our model’s predictions compare to the
true values. We then repeat this process k times.
In the model selection searches described in this chapter (i.e., GridSearchCV and Ran
domizedSearchCV), we used cross-validation to evaluate which hyperparameter values
produced the best models. However, a nuanced and generally underappreciated prob‐
lem arises: since we used the data to select the best hyperparameter values, we cannot
use that same data to evaluate the model’s performance. The solution? Wrap the
cross-validation used for model search in another cross-validation! In nested crossvalidation, the “inner” cross-validation selects the best model, while the “outer” crossvalidation provides us with an unbiased evaluation of the model’s performance. In
our solution, the inner cross-validation is our GridSearchCV object, which we then
wrap in an outer cross-validation using cross_val_score.
If you are confused, try a simple experiment. First, set verbose=1 so we can see what
is happening:
gridsearch = GridSearchCV(logistic, hyperparameters, cv=5, verbose=1)
Next, run gridsearch.fit(features, target), which is our inner cross-validation
used to find the best model:
best_model = gridsearch.fit(features, target)
Fitting 5 folds for each of 20 candidates, totalling 100 fits
[Parallel(n_jobs=1)]: Done 100 out of 100 | elapsed:
0.4s finished
From the output you can see the inner cross-validation trained 20 candidate models
five times, totaling 100 models. Next, nest clf inside a new cross-validation, which
defaults to three folds:
scores = cross_val_score(gridsearch, features, target)
Fitting 5 folds for each of 20 candidates, totalling 100 fits
[Parallel(n_jobs=1)]: Done 100 out of 100 | elapsed:
0.5s finished
12.7 Evaluating Performance After Model Selection
|
221
Fitting 5 folds for each of 20 candidates, totalling 100 fits
[Parallel(n_jobs=1)]: Done 100 out of 100 | elapsed:
0.7s finished
Fitting 5 folds for each of 20 candidates, totalling 100 fits
[Parallel(n_jobs=1)]: Done 100 out of 100 | elapsed:
0.5s finished
The output shows that the inner cross-validation trained 20 models five times to find
the best model, and this model was evaluated using an outer three-fold crossvalidation, creating a total of 300 models trained.
222
|
Chapter 12: Model Selection
CHAPTER 13
Linear Regression
13.0 Introduction
Linear regression is one of the simplest supervised learning algorithms in our toolkit.
If you have ever taken an introductory statistics course in college, likely the final topic
you covered was linear regression. In fact, it is so simple that it is sometimes not con‐
sidered machine learning at all! Whatever you believe, the fact is that linear regres‐
sion—and its extensions—continues to be a common and useful method of making
predictions when the target vector is a quantitative value (e.g., home price, age). In
this chapter we will cover a variety of linear regression methods (and some exten‐
sions) for creating well-performing prediction models.
13.1 Fitting a Line
Problem
You want to train a model that represents a linear relationship between the feature
and target vector.
Solution
Use a linear regression (in scikit-learn, LinearRegression):
# Load libraries
from sklearn.linear_model import LinearRegression
from sklearn.datasets import load_boston
# Load data with only two features
boston = load_boston()
features = boston.data[:,0:2]
target = boston.target
223
# Create linear regression
regression = LinearRegression()
# Fit the linear regression
model = regression.fit(features, target)
Discussion
Linear regression assumes that the relationship between the features and the target
vector is approximately linear. That is, the effect (also called coefficient, weight, or
parameter) of the features on the target vector is constant. In our solution, for the
sake of explanation we have trained our model using only two features. This means
our linear model will be:
y = β 0 + β 1x1 + β 2x2 + �
where ŷ is our target, xi is the data for a single feature, β 1 and β 2 are the coefficients
identified by fitting the model, and ϵ is the error. After we have fit our model, we can
view the value of each parameter. For example, β 0, also called the bias or intercept, can
be viewed using intercept_:
# View the intercept
model.intercept_
22.46681692105723
And β 1 and β 2 are shown using coef_:
# View the feature coefficients
model.coef_
array([-0.34977589,
0.11642402])
In our dataset, the target value is the median value of a Boston home (in the 1970s) in
thousands of dollars. Therefore the price of the first home in the dataset is:
# First value in the target vector multiplied by 1000
target[0]*1000
24000.0
Using the predict method, we can predict a value for that house:
# Predict the target value of the first observation, multiplied by 1000
model.predict(features)[0]*1000
24560.23872370844
Not bad! Our model was only off by $560.24!
224
|
Chapter 13: Linear Regression
The major advantage of linear regression is its interpretability, in large part because
the coefficients of the model are the effect of a one-unit change on the target vector.
For example, the first feature in our solution is the number of crimes per resident.
Our model’s coefficient of this feature was ~–0.35, meaning that if we multiply this
coefficient by 1,000 (since the target vector is the house price in thousands of dollars),
we have the change in house price for each additional one crime per capita:
# First coefficient multiplied by 1000
model.coef_[0]*1000
-349.77588707748947
This says that every single crime per capita will decrease the price of the house by
approximately $350!
13.2 Handling Interactive Effects
Problem
You have a feature whose effect on the target variable depends on another feature.
Solution
Create an interaction term to capture that dependence using scikit-learn’s Polynomial
Features:
# Load libraries
from sklearn.linear_model import LinearRegression
from sklearn.datasets import load_boston
from sklearn.preprocessing import PolynomialFeatures
# Load data with only two features
boston = load_boston()
features = boston.data[:,0:2]
target = boston.target
# Create interaction term
interaction = PolynomialFeatures(
degree=3, include_bias=False, interaction_only=True)
features_interaction = interaction.fit_transform(features)
# Create linear regression
regression = LinearRegression()
# Fit the linear regression
model = regression.fit(features_interaction, target)
13.2 Handling Interactive Effects |
225
Discussion
Sometimes a feature’s effect on our target variable is at least partially dependent on
another feature. For example, imagine a simple coffee-based example where we have
two binary features—the presence of sugar (sugar) and whether or not we have stir‐
red (stirred)—and we want to predict if the coffee tastes sweet. Just putting sugar in
the coffee (sugar=1, stirred=0) won’t make the coffee taste sweet (all the sugar is at
the bottom!) and just stirring the coffee without adding sugar (sugar=0, stirred=1)
won’t make it sweet either. Instead it is the interaction of putting sugar in the coffee
and stirring the coffee (sugar=1, stirred=1) that will make a coffee taste sweet. The
effects of sugar and stir on sweetness are dependent on each other. In this case we
say there is an interaction effect between the features sugar and stirred.
We can account for interaction effects by including a new feature comprising the
product of corresponding values from the interacting features:
y = β 0 + β 1x1 + β 2x2 + β 3x1x2 + �
where x1 and x2 are the values of the sugar and stirred, respectively, and x1x2 repre‐
sents the interaction between the two.
In our solution, we used a dataset containing only two features. Here is the first
observation’s values for each of those features:
# View the feature values for first observation
features[0]
array([
6.32000000e-03,
1.80000000e+01])
To create an interaction term, we simply multiply those two values together for every
observation:
# Import library
import numpy as np
# For each observation, multiply the values of the first and second feature
interaction_term = np.multiply(features[:, 0], features[:, 1])
We can then view the interaction term for the first observation:
# View interaction term for first observation
interaction_term[0]
0.11376
However, while often we will have a substantive reason for believing there is an inter‐
action between two features, sometimes we will not. In those cases it can be useful to
use scikit-learn’s PolynomialFeatures to create interaction terms for all combina‐
226
|
Chapter 13: Linear Regression
tions of features. We can then use model selection strategies to identify the combina‐
tion of features and interaction terms that produce the best model.
To create interaction terms using PolynomialFeatures, there are three important
parameters we must set. Most important, interaction_only=True tells Polynomial
Features to only return interaction terms (and not polynomial features, which we
will discuss in Recipe 13.3). By default, PolynomialFeatures will add a feature con‐
taining ones called a bias. We can prevent that with include_bias=False. Finally, the
degree parameter determines the maximum number of features to create interaction
terms from (in case we wanted to create an interaction term that is the combination
of three features). We can see the output of PolynomialFeatures from our solution
by checking to see if the first observation’s feature values and interaction term value
match our manually calculated version:
# View the values of the first observation
features_interaction[0]
array([
6.32000000e-03,
1.80000000e+01,
1.13760000e-01])
13.3 Fitting a Nonlinear Relationship
Problem
You want to model a nonlinear relationship.
Solution
Create a polynomial regression by including polynomial features in a linear regres‐
sion model:
# Load library
from sklearn.linear_model import LinearRegression
from sklearn.datasets import load_boston
from sklearn.preprocessing import PolynomialFeatures
# Load data with one feature
boston = load_boston()
features = boston.data[:,0:1]
target = boston.target
# Create polynomial features x^2 and x^3
polynomial = PolynomialFeatures(degree=3, include_bias=False)
features_polynomial = polynomial.fit_transform(features)
# Create linear regression
regression = LinearRegression()
# Fit the linear regression
model = regression.fit(features_polynomial, target)
13.3 Fitting a Nonlinear Relationship
|
227
Discussion
So far we have only discussed modeling linear relationships. An example of a linear
relationship would be the number of stories a building has and the building’s height.
In linear regression, we assume the effect of number of stories and building height is
approximately constant, meaning a 20-story building will be roughly twice as high as
a 10-story building, which will be roughly twice as high as a 5-story building. Many
relationships of interest, however, are not strictly linear.
Often we want to model a non-linear relationship—for example, the relationship
between the number of hours a student studies and the score she gets on the test.
Intuitively, we can imagine there is a big difference in test scores between students
who study for one hour compared to students who did not study at all. However,
there is a much smaller difference in test scores between a student who studied for 99
hours and a student who studied for 100 hours. The effect one hour of studying has
on a student’s test score decreases as the number of hours increases.
Polynomial regression is an extension of linear regression to allow us to model non‐
linear relationships. To create a polynomial regression, convert the linear function we
used in Recipe 13.1:
y = β 0 + β 1x1 + �
into a polynomial function by adding polynomial features:
y = β 0 + β 1x1 + β 2x21 + . . . + β dxdi + �
where d is the degree of the polynomial. How are we able to use a linear regression
for a nonlinear function? The answer is that we do not change how the linear regres‐
sion fits the model, but rather only add polynomial features. That is, the linear regres‐
sion does not “know” that the x2 is a quadratic transformation of x. It just considers it
one more variable.
A more practical description might be in order. To model nonlinear relationships, we
can create new features that raise an existing feature, x, up to some power: x2, x3, and
so on. The more of these new features we add, the more flexible the “line” created by
our model. To make this more explicit, imagine we want to create a polynomial to the
third degree. For the sake of simplicity, we will focus on only one observation (the
first observation in the dataset), x0:
# View first observation
features[0]
array([ 0.00632])
228
|
Chapter 13: Linear Regression
To create a polynomial feature, we would raise the first observation’s value to the sec‐
ond degree, x12:
# View first observation raised to the second power, x^2
features[0]**2
array([
3.99424000e-05])
This would be our new feature. We would then also raise the first observation’s value
to the third degree, x13:
# View first observation raised to the third power, x^3
features[0]**3
array([
2.52435968e-07])
By including all three features (x, x2, and x3) in our feature matrix and then running a
linear regression, we have conducted a polynomial regression:
# View the first observation's values for x, x^2, and x^3
features_polynomial[0]
array([
6.32000000e-03,
3.99424000e-05,
2.52435968e-07])
PolynomialFeatures has two important parameters. First, degree determines the
maximum number of degrees for the polynomial features. For example, degree=3 will
generate x2 and x3. Finally, by default PolynomialFeatures includes a feature contain‐
ing only ones (called a bias). We can remove that by setting include_bias=False.
13.4 Reducing Variance with Regularization
Problem
You want to reduce the variance of your linear regression model.
Solution
Use a learning algorithm that includes a shrinkage penalty (also called regularization)
like ridge regression and lasso regression:
# Load libraries
from sklearn.linear_model import Ridge
from sklearn.datasets import load_boston
from sklearn.preprocessing import StandardScaler
# Load data
boston = load_boston()
features = boston.data
target = boston.target
# Standardize features
scaler = StandardScaler()
13.4 Reducing Variance with Regularization
|
229
features_standardized = scaler.fit_transform(features)
# Create ridge regression with an alpha value
regression = Ridge(alpha=0.5)
# Fit the linear regression
model = regression.fit(features_standardized, target)
Discussion
In standard linear regression the model trains to minimize the sum of squared error
between the true (yi) and prediction, (y i) target values, or residual sum of squares
(RSS):
RSS =
n
∑
i=1
yi − y i
2
Regularized regression learners are similar, except they attempt to minimize RSS and
some penalty for the total size of the coefficient values, called a shrinkage penalty
because it attempts to “shrink” the model. There are two common types of regular‐
ized learners for linear regression: ridge regression and the lasso. The only formal dif‐
ference is the type of shrinkage penalty used. In ridge regression, the shrinkage pen‐
alty is a tuning hyperparameter multiplied by the squared sum of all coefficients:
RSS+α
p
2
∑ βj
j=1
where β j is the coefficient of the jth of p features and α is a hyperparameter (discussed
next). The lasso is similar, except the shrinkage penalty is a tuning hyperparameter
multiplied by the sum of the absolute value of all coefficients:
p
1
RSS + α ∑ β j
2n
j=1
where n is the number of observations. So which one should we use? As a very gen‐
eral rule of thumb, ridge regression often produces slightly better predictions than
lasso, but lasso (for reasons we will discuss in Recipe 13.5) produces more interpreta‐
ble models. If we want a balance between ridge and lasso’s penalty functions we can
use elastic net, which is simply a regression model with both penalties included.
Regardless of which one we use, both ridge and lasso regressions can penalize large or
complex models by including coefficient values in the loss function we are trying to
minimize.
230
|
Chapter 13: Linear Regression
The hyperparameter, α, lets us control how much we penalize the coefficients, with
higher values of α creating simpler models. The ideal value of α should be tuned like
any other hyperparameter. In scikit-learn, α is set using the alpha parameter.
scikit-learn includes a RidgeCV method that allows us to select the ideal value for α:
# Load library
from sklearn.linear_model import RidgeCV
# Create ridge regression with three alpha values
regr_cv = RidgeCV(alphas=[0.1, 1.0, 10.0])
# Fit the linear regression
model_cv = regr_cv.fit(features_standardized, target)
# View coefficients
model_cv.coef_
array([-0.91215884,
2.67922108,
-2.05390717,
1.0658758 , 0.11942614, 0.68558782, -2.03231631,
0.01477326, -3.0777265 , 2.58814315, -2.00973173,
0.85614763, -3.73565106])
We can then easily view the best model’s α value:
# View alpha
model_cv.alpha_
1.0
One final note: because in linear regression the value of the coefficients is partially
determined by the scale of the feature, and in regularized models all coefficients are
summed together, we must make sure to standardize the feature prior to training.
13.5 Reducing Features with Lasso Regression
Problem
You want to simplify your linear regression model by reducing the number of fea‐
tures.
Solution
Use a lasso regression:
# Load library
from sklearn.linear_model import Lasso
from sklearn.datasets import load_boston
from sklearn.preprocessing import StandardScaler
# Load data
boston = load_boston()
13.5 Reducing Features with Lasso Regression
|
231
features = boston.data
target = boston.target
# Standardize features
scaler = StandardScaler()
features_standardized = scaler.fit_transform(features)
# Create lasso regression with alpha value
regression = Lasso(alpha=0.5)
# Fit the linear regression
model = regression.fit(features_standardized, target)
Discussion
One interesting characteristic of lasso regression’s penalty is that it can shrink the
coefficients of a model to zero, effectively reducing the number of features in the
model. For example, in our solution we set alpha to 0.5 and we can see that many of
the coefficients are 0, meaning their corresponding features are not used in the
model:
# View coefficients
model.coef_
array([-0.10697735, 0.
, -0.
, 0.39739898, -0.
2.97332316, -0.
, -0.16937793, -0.
, -0.
-1.59957374, 0.54571511, -3.66888402])
,
,
However, if we increase α to a much higher value, we see that literally none of the
features are being used:
# Create lasso regression with a high alpha
regression_a10 = Lasso(alpha=10)
model_a10 = regression_a10.fit(features_standardized, target)
model_a10.coef_
array([-0.,
0., -0.,
0., -0.,
0., -0.,
0., -0., -0., -0.,
0., -0.])
The practical benefit of this effect is that it means that we could include 100 features
in our feature matrix and then, through adjusting lasso’s α hyperparameter, produce a
model that uses only 10 (for instance) of the most important features. This lets us
reduce variance while improving the interpretability of our model (since fewer fea‐
tures is easier to explain).
232
|
Chapter 13: Linear Regression
CHAPTER 14
Trees and Forests
14.0 Introduction
Tree-based learning algorithms are a broad and popular family of related nonparametric, supervised methods for both classification and regression. The basis of
tree-based learners is the decision tree wherein a series of decision rules (e.g., “If their
gender is male…”) are chained. The result looks vaguely like an upside-down tree,
with the first decision rule at the top and subsequent decision rules spreading out
below. In a decision tree, every decision rule occurs at a decision node, with the rule
creating branches leading to new nodes. A branch without a decision rule at the end
is called a leaf.
One reason for the popularity of tree-based models is their interpretability. In fact,
decision trees can literally be drawn out in their complete form (see Recipe 14.3) to
create a highly intuitive model. From this basic tree system comes a wide variety of
extensions from random forests to stacking. In this chapter we will cover how to
train, handle, adjust, visualize, and evaluate a number of tree-based models.
14.1 Training a Decision Tree Classifier
Problem
You need to train a classifier using a decision tree.
Solution
Use scikit-learn’s DecisionTreeClassifier:
# Load libraries
from sklearn.tree import DecisionTreeClassifier
233
from sklearn import datasets
# Load data
iris = datasets.load_iris()
features = iris.data
target = iris.target
# Create decision tree classifier object
decisiontree = DecisionTreeClassifier(random_state=0)
# Train model
model = decisiontree.fit(features, target)
Discussion
Decision tree learners attempt to find a decision rule that produces the greatest
decrease in impurity at a node. While there are a number of measurements of impur‐
ity, by default DecisionTreeClassifier uses Gini impurity:
G t =1−
c
∑ pi2
i=1
where G(t) is the Gini impurity at node t and pi is the proportion of observations of
class c at node t. This process of finding the decision rules that create splits to increase
impurity is repeated recursively until all leaf nodes are pure (i.e., contain only one
class) or some arbitrary cut-off is reached.
In scikit-learn, DecisionTreeClassifier operates like other learning methods; after
the model is trained using fit we can use the model to predict the class of an obser‐
vation:
# Make new observation
observation = [[ 5, 4, 3, 2]]
# Predict observation's class
model.predict(observation)
array([1])
We can also see the predicted class probabilities of the observation:
# View predicted class probabilities for the three classes
model.predict_proba(observation)
array([[ 0.,
1.,
0.]])
Finally, if we want to use a different impurity measurement we can use the criterion
parameter:
# Create decision tree classifier object using entropy
decisiontree_entropy = DecisionTreeClassifier(
234
|
Chapter 14: Trees and Forests
criterion='entropy', random_state=0)
# Train model
model_entropy = decisiontree_entropy.fit(features, target)
See Also
• Decision Tree Learning, Princeton
14.2 Training a Decision Tree Regressor
Problem
You need to train a regression model using a decision tree.
Solution
Use scikit-learn’s DecisionTreeRegressor:
# Load libraries
from sklearn.tree import DecisionTreeRegressor
from sklearn import datasets
# Load data with only two features
boston = datasets.load_boston()
features = boston.data[:,0:2]
target = boston.target
# Create decision tree classifier object
decisiontree = DecisionTreeRegressor(random_state=0)
# Train model
model = decisiontree.fit(features, target)
Discussion
Decision tree regression works similarly to decision tree classification; however,
instead of reducing Gini impurity or entropy, potential splits are by default measured
on how much they reduce mean squared error (MSE):
MSE =
1 n
yi − y i
ni∑
=1
2
where yi is the true value of the target and y i is the predicted value. In scikit-learn,
decision tree regression can be conducted using DecisionTreeRegressor. Once we
14.2 Training a Decision Tree Regressor
|
235
have trained a decision tree, we can use it to predict the target value for an observa‐
tion:
# Make new observation
observation = [[0.02, 16]]
# Predict observation's value
model.predict(observation)
array([ 33.])
Just like with DecisionTreeClassifier we can use the criterion parameter to select
the desired measurement of split quality. For example, we can construct a tree whose
splits reduce mean absolute error (MAE):
# Create decision tree classifier object using entropy
decisiontree_mae = DecisionTreeRegressor(criterion="mae", random_state=0)
# Train model
model_mae = decisiontree_mae.fit(features, target)
See Also
• Decision Tree Regression, scikit-learn
14.3 Visualizing a Decision Tree Model
Problem
You need to visualize a model created by a decision tree learning algorithm.
Solution
Export the decision tree model into DOT format, then visualize:
# Load libraries
import pydotplus
from sklearn.tree import DecisionTreeClassifier
from sklearn import datasets
from IPython.display import Image
from sklearn import tree
# Load data
iris = datasets.load_iris()
features = iris.data
target = iris.target
# Create decision tree classifier object
decisiontree = DecisionTreeClassifier(random_state=0)
236
|
Chapter 14: Trees and Forests
# Train model
model = decisiontree.fit(features, target)
# Create DOT data
dot_data = tree.export_graphviz(decisiontree,
out_file=None,
feature_names=iris.feature_names,
class_names=iris.target_names)
# Draw graph
graph = pydotplus.graph_from_dot_data(dot_data)
# Show graph
Image(graph.create_png())
Discussion
One of the advantages of decision tree classifiers is that we can visualize the entire
trained model—making decision trees one of the most interpretable models in
machine learning. In our solution, we exported our trained model in DOT format (a
graph description language) and then used that to draw the graph.
14.3 Visualizing a Decision Tree Model
|
237
If we look at the root node, we can see the decision rule is that if petal widths are less
than or equal to 0.8, then go to the left branch; if not, go to the right branch. We can
also see the Gini impurity index (0.667), the number of observations (150), the num‐
ber of observations in each class ([50,50,50]), and the class the observations would
be predicted to be if we stopped at that node (setosa). We can also see that at that
node the learner found that a single decision rule (petal width (cm) <= 0.8) was
able to perfectly identify all of the setosa class observations. Furthermore, with one
more decision rule with the same feature (petal width (cm) <= 1.75) the decision
tree is able to correctly classify 144 of 150 observations. This makes petal width a very
important feature!
If we want to use the decision tree in other applications or reports, we can easily
export the visualization into PDF or a PNG image:
# Create PDF
graph.write_pdf("iris.pdf")
True
# Create PNG
graph.write_png("iris.png")
True
While this solution visualized a decision tree classifier, it can just as easily be used to
visualize a decision tree regressor.
Note: macOS users might have to install GraphViz’s executable to run the preceding
code. This can be done using Homebrew: brew install graphviz. For Homebrew
installation instructions, visit Homebrew’s website.
See Also
• Homebrew
14.4 Training a Random Forest Classifier
Problem
You want to train a classification model using a “forest” of randomized decision trees.
Solution
Train a random forest classification model using scikit-learn’s RandomForestClassi
fier:
238
|
Chapter 14: Trees and Forests
# Load libraries
from sklearn.ensemble import RandomForestClassifier
from sklearn import datasets
# Load data
iris = datasets.load_iris()
features = iris.data
target = iris.target
# Create random forest classifier object
randomforest = RandomForestClassifier(random_state=0, n_jobs=-1)
# Train model
model = randomforest.fit(features, target)
Discussion
A common problem with decision trees is that they tend to fit the training data too
closely (i.e., overfitting). This has motivated the widespread use of an ensemble learn‐
ing method called random forest. In a random forest, many decision trees are trained,
but each tree only receives a bootstrapped sample of observations (i.e., a random
sample of observations with replacement that matches the original number of obser‐
vations) and each node only considers a subset of features when determining the best
split. This forest of randomized decision trees (hence the name) votes to determine
the predicted class.
As we can see by comparing this solution to Recipe 14.1, scikit-learn’s RandomForest
Classifier works similarly to DecisionTreeClassifier:
# Make new observation
observation = [[ 5, 4, 3, 2]]
# Predict observation's class
model.predict(observation)
array([1])
RandomForestClassifier also uses many of the same parameters as DecisionTree
Classifier. For example, we can change the measure of split quality used:
# Create random forest classifier object using entropy
randomforest_entropy = RandomForestClassifier(
criterion="entropy", random_state=0)
# Train model
model_entropy = randomforest_entropy.fit(features, target)
However, being a forest rather than an individual decision tree, RandomForestClassi
fier has certain parameters that are either unique to random forests or particularly
important. First, the max_features parameter determines the maximum number of
14.4 Training a Random Forest Classifier
|
239
features to be considered at each node and takes a number of arguments including
integers (number of features), floats (percentage of features), and sqrt (square root of
the number of features). By default, max_features is set to auto, which acts the same
as sqrt. Second, the bootstrap parameter allows us to set whether the subset of
observations considered for a tree is created using sampling with replacement (the
default setting) or without replacement. Third, n_estimators sets the number of
decision trees to include in the forest. In Recipe 10.4 we treated n_estimators as a
hyperparameter and visualized the effect of increasing the number of trees on an
evaluation metric. Finally, while not specific to random forest classifiers, because we
are effectively training many decision tree models, it is often useful to use all available
cores by setting n_jobs=-1.
See Also
• Random Forests, Berkeley Statistics
14.5 Training a Random Forest Regressor
Problem
You want to train a regression model using a “forest” of randomized decision trees.
Solution
Train a random forest regression model using scikit-learn’s RandomForestRegressor:
# Load libraries
from sklearn.ensemble import RandomForestRegressor
from sklearn import datasets
# Load data with only two features
boston = datasets.load_boston()
features = boston.data[:,0:2]
target = boston.target
# Create random forest classifier object
randomforest = RandomForestRegressor(random_state=0, n_jobs=-1)
# Train model
model = randomforest.fit(features, target)
Discussion
Just like how we can make a forest of decision tree classifiers, we can make a forest of
decision tree regressors where each tree uses a bootstrapped subset of observations
240
|
Chapter 14: Trees and Forests
and at each node the decision rule considers only a subset of features. As with Random
ForestClassifier we have certain important parameters:
• max_features sets the maximum number of features to consider at each node.
Defaults to p features, where p is the total number of features.
• bootstrap sets whether or not to sample with replacement. Defaults to True.
• n_estimators sets the number of decision trees to construct. Defaults to 10.
See Also
• RandomForestRegressor, scikit-learn
14.6 Identifying Important Features in Random Forests
Problem
You need to know which features are most important in a random forest model.
Solution
Calculate and visualize the importance of each feature:
# Load libraries
import numpy as np
import matplotlib.pyplot as plt
from sklearn.ensemble import RandomForestClassifier
from sklearn import datasets
# Load data
iris = datasets.load_iris()
features = iris.data
target = iris.target
# Create random forest classifier object
randomforest = RandomForestClassifier(random_state=0, n_jobs=-1)
# Train model
model = randomforest.fit(features, target)
# Calculate feature importances
importances = model.feature_importances_
# Sort feature importances in descending order
indices = np.argsort(importances)[::-1]
# Rearrange feature names so they match the sorted feature importances
14.6 Identifying Important Features in Random Forests
|
241
names = [iris.feature_names[i] for i in indices]
# Create plot
plt.figure()
# Create plot title
plt.title("Feature Importance")
# Add bars
plt.bar(range(features.shape[1]), importances[indices])
# Add feature names as x-axis labels
plt.xticks(range(features.shape[1]), names, rotation=90)
# Show plot
plt.show()
242
|
Chapter 14: Trees and Forests
Discussion
One of the major benefits of decision trees is interpretability. Specifically, we can visu‐
alize the entire model (see Recipe 13.3). However, a random forest model is com‐
prised of tens, hundreds, even thousands of decision trees. This makes a simple, intu‐
itive visualization of a random forest model impractical. That said, there is another
option: we can compare (and visualize) the relative importance of each feature.
In Recipe 13.3, we visualized a decision tree classifier model and saw that decision
rules based only on petal width were able to classify many observations correctly.
Intuitively, we can say that this means that petal width is an important feature in our
classifier. More formally, features with splits that have the greater mean decrease in
impurity (e.g., Gini impurity or entropy in classifiers and variance in regressors) are
considered more important.
However, there are two things to keep in mind regarding feature importance. First,
scikit-learn requires that we break up nominal categorical features into multiple
binary features. This has the effect of spreading the importance of that feature across
all of the binary features and can often make each feature appear to be unimportant
even when the original nominal categorical feature is highly important. Second, if two
features are highly correlated, one feature will claim much of the importance, making
the other feature appear to be far less important—which has implications for inter‐
pretation if not considered.
In scikit-learn, classification and regression decision trees and random forests can
report the relative importance of each feature using the feature_importances_
method:
# View feature importances
model.feature_importances_
array([ 0.11896532,
0.0231668 ,
0.36804744,
0.48982043])
The higher the number, the more important the feature (all importance scores sum to
1). By plotting these values we can add interpretability to our random forest models.
14.7 Selecting Important Features in Random Forests
Problem
You need to conduct feature selection on a random forest.
Solution
Identify the importance features and retrain the model using only the most important
features:
14.7 Selecting Important Features in Random Forests
|
243
# Load libraries
from sklearn.ensemble import RandomForestClassifier
from sklearn import datasets
from sklearn.feature_selection import SelectFromModel
# Load data
iris = datasets.load_iris()
features = iris.data
target = iris.target
# Create random forest classifier
randomforest = RandomForestClassifier(random_state=0, n_jobs=-1)
# Create object that selects features with importance greater
# than or equal to a threshold
selector = SelectFromModel(randomforest, threshold=0.3)
# Feature new feature matrix using selector
features_important = selector.fit_transform(features, target)
# Train random forest using most important featres
model = randomforest.fit(features_important, target)
Discussion
There are situations where we might want to reduce the number of features in our
model. For example, we might want to reduce the model’s variance or we might want
to improve interpretability by including only the most important features.
In scikit-learn we can use a simple two-stage workflow to create a model with
reduced features. First, we train a random forest model using all features. Then, we
use this model to identify the most important features. Next, we create a new feature
matrix that includes only these features. In our solution, we used the SelectFromMo
del method to create a feature matrix containing only features with an importance
greater than or equal to some threshold value. Finally, we created a new model using
only those features.
It must be noted that there are two caveats to this approach. First, nominal categorical
features that have been one-hot encoded will see the feature importance diluted
across the binary features. Second, the feature importance of highly correlated fea‐
tures will be effectively assigned to one feature and not evenly distributed across both
features.
See Also
• Variable selection using Random Forests, Robin Genuer, Jean-Michel Poggi,
Christine Tuleau-Malot
244
|
Chapter 14: Trees and Forests
14.8 Handling Imbalanced Classes
Problem
You have a target vector with highly imbalanced classes and want to train a random
forest model.
Solution
Train a decision tree or random forest model with class_weight="balanced":
# Load libraries
import numpy as np
from sklearn.ensemble import RandomForestClassifier
from sklearn import datasets
# Load data
iris = datasets.load_iris()
features = iris.data
target = iris.target
# Make class highly imbalanced by removing first 40 observations
features = features[40:,:]
target = target[40:]
# Create target vector indicating if class 0, otherwise 1
target = np.where((target == 0), 0, 1)
# Create random forest classifier object
randomforest = RandomForestClassifier(
random_state=0, n_jobs=-1, class_weight="balanced")
# Train model
model = randomforest.fit(features, target)
Discussion
Imbalanced classes are a common problem when we are doing machine learning in
the real world. Left unaddressed, the presence of imbalanced classes can reduce the
performance of our model. We already talked about a few strategies for handling
imbalanced classes during preprocessing in Recipe 17.5. However, many learning
algorithms in scikit-learn come with built-in methods for correcting for imbalanced
classes. We can set RandomForestClassifier to correct for imbalanced classes using
the class_weight parameter. If supplied with a dictionary in the form of class names
and respective desired weights (e.g., {"male": 0.2, "female": 0.8}), RandomFor
estClassifier will weight the classes accordingly. However, often a more useful
argument is balanced, wherein classes are automatically weighted inversely propor‐
tional to how frequently they appear in the data:
14.8 Handling Imbalanced Classes
|
245
wj =
n
kn j
where wj is the weight to class j, n is the number of observations, nj is the number of
observations in class j, and k is the total number of classes. For example, in our solu‐
tion we have 2 classes (k), 110 observations (n), and 10 and 100 observations in each
class, respectively (nj). If we weight the classes using class_weight="balanced", then
the smaller class is weighted more:
# Calculate weight for small class
110/(2*10)
5.5
while the larger class is weighted less:
# Calculate weight for large class
110/(2*100)
0.55
14.9 Controlling Tree Size
Problem
You want to manually determine the structure and size of a decision tree.
Solution
Use the tree structure parameters in scikit-learn tree-based learning algorithms:
# Load libraries
from sklearn.tree import DecisionTreeClassifier
from sklearn import datasets
# Load data
iris = datasets.load_iris()
features = iris.data
target = iris.target
# Create decision tree classifier object
decisiontree = DecisionTreeClassifier(random_state=0,
max_depth=None,
min_samples_split=2,
min_samples_leaf=1,
min_weight_fraction_leaf=0,
max_leaf_nodes=None,
min_impurity_decrease=0)
246
|
Chapter 14: Trees and Forests
# Train model
model = decisiontree.fit(features, target)
Discussion
scikit-learn’s tree-based learning algorithms have a variety of techniques for control‐
ling the size of the decision tree(s). These are accessed through parameters:
max_depth
Maximum depth of the tree. If None, the tree is grown until all leaves are pure. If
an integer, the tree is effectively “pruned” to that depth.
min_samples_split
Minimum number of observations at a node before that node is split. If an inte‐
ger is supplied as an argument it determines the raw minimum, while if a float is
supplied the minimum is the percent of total observations.
min_samples_leaf
Minimum number of observations required to be at a leaf. Uses the same argu‐
ments as min_samples_split.
max_leaf_nodes
Maximum number of leaves.
min_impurity_split
Minimum impurity decrease required before a split is performed.
While it is useful to know these parameters exist, most likely we will only be using
max_depth and min_impurity_split because shallower trees (sometimes called
stumps) are simpler models and thus have lower variance.
14.10 Improving Performance Through Boosting
Problem
You need a model with better performance than decision trees or random forests.
Solution
Train a boosted model using AdaBoostClassifier or AdaBoostRegressor:
# Load libraries
from sklearn.ensemble import AdaBoostClassifier
from sklearn import datasets
# Load data
iris = datasets.load_iris()
features = iris.data
14.10 Improving Performance Through Boosting
|
247
target = iris.target
# Create adaboost tree classifier object
adaboost = AdaBoostClassifier(random_state=0)
# Train model
model = adaboost.fit(features, target)
Discussion
In random forest, an ensemble (group) of randomized decision trees predicts the tar‐
get vector. An alternative, and often more powerful, approach is called boosting. In
one form of boosting called AdaBoost, we iteratively train a series of weak models
(most often a shallow decision tree, sometimes called a stump), each iteration giving
higher priority to observations the previous model predicted incorrectly. More specif‐
ically, in AdaBoost:
1
1. Assign every observation, xi, an initial weight value, wi = n , where n is the total
number of observations in the data.
2. Train a “weak” model on the data.
3. For each observation:
a. If weak model predicts xi correctly, wi is increased.
b. If weak model predicts xi incorrectly, wi is decreased.
4. Train a new weak model where observations with greater wi are given greater pri‐
ority.
5. Repeat steps 4 and 5 until the data is perfectly predicted or a preset number of
weak models has been trained.
The end result is an aggregated model where individual weak models focus on more
difficult (from a prediction perspective) observations. In scikit-learn, we can imple‐
ment AdaBoost using AdaBoostClassifier or AdaBoostRegressor. The most impor‐
tant parameters are base_estimator, n_estimators, and learning_rate:
• base_estimator is the learning algorithm to use to train the weak models. This
will almost always not need to be changed because by far the most common
learner to use with AdaBoost is a decision tree—the parameter’s default argu‐
ment.
• n_estimators is the number of models to iteratively train.
• learning_rate is the contribution of each model to the weights and defaults to 1.
Reducing the learning rate will mean the weights will be increased or decreased
248
|
Chapter 14: Trees and Forests
to a small degree, forcing the model to train slower (but sometimes resulting in
better performance scores).
• loss is exclusive to AdaBoostRegressor and sets the loss function to use when
updating weights. This defaults to a linear loss function, but can be changed to
square or exponential.
See Also
• Explaining AdaBoost, Robert E. Schapire
14.11 Evaluating Random Forests with Out-of-Bag Errors
Problem
You need to evaluate a random forest model without using cross-validation.
Solution
Calculate the model’s out-of-bag score:
# Load libraries
from sklearn.ensemble import RandomForestClassifier
from sklearn import datasets
# Load data
iris = datasets.load_iris()
features = iris.data
target = iris.target
# Create random tree classifier object
randomforest = RandomForestClassifier(
random_state=0, n_estimators=1000, oob_score=True, n_jobs=-1)
# Train model
model = randomforest.fit(features, target)
# View out-of-bag-error
randomforest.oob_score_
0.95333333333333337
Discussion
In random forests, each decision tree is trained using a bootstrapped subset of obser‐
vations. This means that for every tree there is a separate subset of observations not
14.11 Evaluating Random Forests with Out-of-Bag Errors
|
249
being used to train that tree. These are called out-of-bag (OOB) observations. We can
use OOB observations as a test set to evaluate the performance of our random forest.
For every observation, the learning algorithm compares the observation’s true value
with the prediction from a subset of trees not trained using that observation. The
overall score is calculated and provides a single measure of a random forest’s perfor‐
mance. OOB score estimation is an alternative to cross-validation.
In scikit-learn, we can OOB scores of a random forest by setting oob_score=True in
the random forest object (i.e., RandomForestClassifier). The score can be retrieved
using oob_score_.
250
|
Chapter 14: Trees and Forests
CHAPTER 15
K-Nearest Neighbors
15.0 Introduction
The K-Nearest Neighbors classifier (KNN) is one of the simplest yet most commonly
used classifiers in supervised machine learning. KNN is often considered a lazy
learner; it doesn’t technically train a model to make predictions. Instead an observa‐
tion is predicted to be the class of that of the largest proportion of the k nearest obser‐
vations. For example, if an observation with an unknown class is surrounded by an
observation of class 1, then the observation is classified as class 1. In this chapter we
will explore how to use scikit-learn to create and use a KNN classifier.
15.1 Finding an Observation’s Nearest Neighbors
Problem
You need to find an observation’s k nearest observations (neighbors).
Solution
Use scikit-learn’s NearestNeighbors:
# Load libraries
from sklearn import datasets
from sklearn.neighbors import NearestNeighbors
from sklearn.preprocessing import StandardScaler
# Load data
iris = datasets.load_iris()
features = iris.data
# Create standardizer
251
standardizer = StandardScaler()
# Standardize features
features_standardized = standardizer.fit_transform(features)
# Two nearest neighbors
nearest_neighbors = NearestNeighbors(n_neighbors=2).fit(features_standardized)
# Create an observation
new_observation = [ 1, 1,
1,
1]
# Find distances and indices of the observation's nearest neighbors
distances, indices = nearest_neighbors.kneighbors([new_observation])
# View the nearest neighbors
features_standardized[indices]
array([[[ 1.03800476,
[ 0.79566902,
0.56925129,
0.33784833,
1.10395287,
0.76275864,
1.1850097 ],
1.05353673]]])
Discussion
In our solution we used the dataset of Iris flowers. We created an observation,
new_observation, with some values and then found the two observations that are
closest to our observation. indices contains the locations of the observations in our
dataset that are closest, so X[indices] displays the values of those observations.
Intuitively, distance can be thought of as a measure of similarity, so the two closest
observations are the two flowers most similar to the flower we created.
How do we measure distance? scikit-learn offers a wide variety of distance metrics, d,
including Euclidean:
deuclidean =
∑ni = 1
xi − yi
2
and Manhattan distance:
dmanhattan =
n
∑ xi − yi
i=1
By default, NearestNeighbors uses Minkowski distance:
dminkowski =
252
|
n
∑ xi − yi p
i=1
1/ p
Chapter 15: K-Nearest Neighbors
where xi and yi are the two observations we are calculating the distance between.
Minkowski includes a hyperparameter, p, where p = 1 is Manhattan distance and p =
2 is Euclidean distance, and so on. By default in scikit-learn p = 2.
We can set the distance metric using the metric parameter:
# Find two nearest neighbors based on euclidean distance
nearestneighbors_euclidean = NearestNeighbors(
n_neighbors=2, metric='euclidean').fit(features_standardized)
The distance variable we created contains the actual distance measurement to each
of the two nearest neighbors:
# View distances
distances
array([[ 0.48168828,
0.73440155]])
In addition, we can use kneighbors_graph to create a matrix indicating each observa‐
tion’s nearest neighbors:
# Find each observation's three nearest neighbors
# based on euclidean distance (including itself)
nearestneighbors_euclidean = NearestNeighbors(
n_neighbors=3, metric="euclidean").fit(features_standardized)
# List of lists indicating each observation's 3 nearest neighbors
# (including itself)
nearest_neighbors_with_self = nearestneighbors_euclidean.kneighbors_graph(
features_standardized).toarray()
# Remove 1's marking an observation is a nearest neighbor to itself
for i, x in enumerate(nearest_neighbors_with_self):
x[i] = 0
# View first observation's two nearest neighbors
nearest_neighbors_with_self[0]
array([ 0.,
0.,
0.,
0.,
0.,
0.,
0.,
0.,
0.,
0.,
0.,
0.,
0.,
0.,
1.,
0.,
0.,
0.,
0.,
0.,
0.,
0.,
0.,
0.,
0.,
0.,
0.,
0.,
0.,
0.,
0.,
0.,
0.,
0.,
0.,
0.,
0.,
0.,
0.,
0.,
0.,
0.,
0.,
0.,
0.,
0.,
0.,
0.,
0.,
1.,
0.,
0.,
0.,
0.,
0.,
0.,
0.,
0.,
0.,
0.,
0.,
0.,
0.,
0.,
0.,
0.,
0.,
0.,
0.,
0.,
0.,
0.,
0.,
0.,
0.,
0.,
0.,
0.,
0.,
0.,
0.,
0.,
0.,
0.])
0.,
0.,
0.,
0.,
0.,
0.,
0.,
0.,
0.,
0.,
0.,
0.,
0.,
0.,
0.,
0.,
0.,
0.,
0.,
0.,
0.,
0.,
0.,
0.,
0.,
0.,
0.,
0.,
0.,
0.,
0.,
0.,
0.,
0.,
0.,
0.,
0.,
0.,
0.,
0.,
0.,
0.,
0.,
0.,
0.,
0.,
0.,
0.,
0.,
0.,
0.,
0.,
0.,
0.,
0.,
0.,
0.,
0.,
0.,
0.,
0.,
0.,
0.,
0.,
0.,
0.,
When we are finding nearest neighbors or using any learning algorithm based on dis‐
tance, it is important to transform features so that they are on the same scale. The
15.1 Finding an Observation’s Nearest Neighbors
|
253
reason is because the distance metrics treat all features as if they were on the same
scale, but if one feature is in millions of dollars and a second feature is in percentages,
the distance calculated will be biased toward the former. In our solution we addressed
this potential issue by standardizing the features using StandardScaler.
15.2 Creating a K-Nearest Neighbor Classifier
Problem
Given an observation of unknown class, you need to predict its class based on the
class of its neighbors.
Solution
If the dataset is not very large, use KNeighborsClassifier:
# Load libraries
from sklearn.neighbors import KNeighborsClassifier
from sklearn.preprocessing import StandardScaler
from sklearn import datasets
# Load data
iris = datasets.load_iris()
X = iris.data
y = iris.target
# Create standardizer
standardizer = StandardScaler()
# Standardize features
X_std = standardizer.fit_transform(X)
# Train a KNN classifier with 5 neighbors
knn = KNeighborsClassifier(n_neighbors=5, n_jobs=-1).fit(X_std, y)
# Create two observations
new_observations = [[ 0.75, 0.75, 0.75,
[ 1, 1, 1, 1]]
0.75],
# Predict the class of two observations
knn.predict(new_observations)
array([1, 2])
Discussion
In KNN, given an observation, xu, with an unknown target class, the algorithm first
identifies the k closest observations (sometimes called xu’s neighborhood) based on
some distance metric (e.g., Euclidean distance), then these k observations “vote”
254
|
Chapter 15: K-Nearest Neighbors
based on their class, and the class that wins the vote is xu’s predicted class. More for‐
mally, the probability xu is some class j is:
1
I yi = j
ki∑
∈ν
where ν is the k observation in xu’s neighborhood, yi is the class of the ith observation,
and I is an indicator function (i.e., 1 is true, 0 otherwise). In scikit-learn we can see
these probabilities using predict_proba:
# View probability each observation is one of three classes
knn.predict_proba(new_observations)
array([[ 0. ,
[ 0. ,
0.6,
0. ,
0.4],
1. ]])
The class with the highest probability becomes the predicted class. For example, in
the preceding output, the first observation should be class 1 (Pr = 0.6) while the sec‐
ond observation should be class 2 (Pr = 1), and this is just what we see:
knn.predict(new_observations)
array([1, 2])
KNeighborsClassifier contains a number of important parameters to consider.
First, metric sets the distance metric used (discussed more in Recipe 14.1). Second,
n_jobs determines how many of the computer’s cores to use. Because making a pre‐
diction requires calculating the distance from a point to every single point in the data,
using multiple cores is highly recommended. Third, algorithm sets the method used
to calculate the nearest neighbors. While there are real differences in the algorithms,
by default KNeighborsClassifier attempts to auto-select the best algorithm so you
often don’t need to worry about this parameter. Fourth, by default KNeighborsClassi
fier works how we described previously, with each observation in the neighborhood
getting one vote; however, if we set the weights parameter to distance, the closer
observations’ votes are weighted more than observations farther away. Intuitively this
make sense, since more similar neighbors might tell us more about an observation’s
class than others.
Finally, as discussed in Recipe 14.1, because distance calculations treat all features as if
they are on the same scale, it is important to standardize the features prior to using a
KNN classifier.
15.2 Creating a K-Nearest Neighbor Classifier
|
255
15.3 Identifying the Best Neighborhood Size
Problem
You want to select the best value for k in a k-nearest neighbors classifier.
Solution
Use model selection techniques like GridSearchCV:
# Load libraries
from sklearn.neighbors import KNeighborsClassifier
from sklearn import datasets
from sklearn.preprocessing import StandardScaler
from sklearn.pipeline import Pipeline, FeatureUnion
from sklearn.model_selection import GridSearchCV
# Load data
iris = datasets.load_iris()
features = iris.data
target = iris.target
# Create standardizer
standardizer = StandardScaler()
# Standardize features
features_standardized = standardizer.fit_transform(features)
# Create a KNN classifier
knn = KNeighborsClassifier(n_neighbors=5, n_jobs=-1)
# Create a pipeline
pipe = Pipeline([("standardizer", standardizer), ("knn", knn)])
# Create space of candidate values
search_space = [{"knn__n_neighbors": [1, 2, 3, 4, 5, 6, 7, 8, 9, 10]}]
# Create grid search
classifier = GridSearchCV(
pipe, search_space, cv=5, verbose=0).fit(features_standardized, target)
Discussion
The size of k has real implications in KNN classifiers. In machine learning we are try‐
ing to find a balance between bias and variance, and in few places is that as explicit as
the value of k. If k = n where n is the number of observations, then we have high bias
but low variance. If k = 1, we will have low bias but high variance. The best model will
come from finding the value of k that balances this bias-variance trade-off. In our sol‐
ution, we used GridSearchCV to conduct five-fold cross-validation on KNN classifiers
256
|
Chapter 15: K-Nearest Neighbors
with different values of k. When that is completed, we can see the k that produces the
best model:
# Best neighborhood size (k)
classifier.best_estimator_.get_params()["knn__n_neighbors"]
6
15.4 Creating a Radius-Based Nearest Neighbor Classifier
Problem
Given an observation of unknown class, you need to predict its class based on the
class of all observations within a certain distance.
Solution
Use RadiusNeighborsClassifier:
# Load libraries
from sklearn.neighbors import RadiusNeighborsClassifier
from sklearn.preprocessing import StandardScaler
from sklearn import datasets
# Load data
iris = datasets.load_iris()
features = iris.data
target = iris.target
# Create standardizer
standardizer = StandardScaler()
# Standardize features
features_standardized = standardizer.fit_transform(features)
# Train a radius neighbors classifier
rnn = RadiusNeighborsClassifier(
radius=.5, n_jobs=-1).fit(features_standardized, target)
# Create two observations
new_observations = [[ 1, 1,
1,
1]]
# Predict the class of two observations
rnn.predict(new_observations)
array([2])
Discussion
In KNN classification, an observation’s class is predicted from the classes of its k
neighbors. A less common technique is classification in a radius-based (RNN) classi‐
15.4 Creating a Radius-Based Nearest Neighbor Classifier
|
257
fier, where an observation’s class is predicted from the classes of all observations
within a given radius r. In scikit-learn RadiusNeighborsClassifier is very similar to
KNeighborsClassifier, with the exception of two parameters. First, in RadiusNeigh
borsClassifier we need to specify the radius of the fixed area used to determine if
an observation is a neighbor using radius. Unless there is some substantive reason
for setting radius to some value, it is best to treat it like any other hyperparameter
and tune it during model selection. The second useful parameter is outlier_label,
which indicates what label to give an observation that has no observations within the
radius—which itself can often be a useful tool for identifying outliers.
258
|
Chapter 15: K-Nearest Neighbors
CHAPTER 16
Logistic Regression
16.0 Introduction
Despite being called a regression, logistic regression is actually a widely used super‐
vised classification technique. Logistic regression and its extensions, like multinomial
logistic regression, allow us to predict the probability that an observation is of a cer‐
tain class using a straightforward and well-understood approach. In this chapter, we
will cover training a variety of classifiers using scikit-learn.
16.1 Training a Binary Classifier
Problem
You need to train a simple classifier model.
Solution
Train a logistic regression in scikit-learn using LogisticRegression:
# Load libraries
from sklearn.linear_model import LogisticRegression
from sklearn import datasets
from sklearn.preprocessing import StandardScaler
# Load data with only two classes
iris = datasets.load_iris()
features = iris.data[:100,:]
target = iris.target[:100]
# Standardize features
scaler = StandardScaler()
features_standardized = scaler.fit_transform(features)
259
# Create logistic regression object
logistic_regression = LogisticRegression(random_state=0)
# Train model
model = logistic_regression.fit(features_standardized, target)
Discussion
Despite having “regression” in its name, a logistic regression is actually a widely used
binary classifier (i.e., the target vector can only take two values). In a logistic regres‐
sion, a linear model (e.g., β0 + β1x) is included in a logistic (also called sigmoid) func‐
1
tion,
, such that:
−z
1+e
P yi = 1 ∣ X =
1
− β0 + β1x
1+e
where P(yi = 1 | X) is the probability of the ith observation’s target value, yi, being class
1, X is the training data, β0 and β1 are the parameters to be learned, and e is Euler’s
number. The effect of the logistic function is to constrain the value of the function’s
output to between 0 and 1 so that it can be interpreted as a probability. If P(yi = 1 | X)
is greater than 0.5, class 1 is predicted; otherwise, class 0 is predicted.
In scikit-learn, we can learn a logistic regression model using LogisticRegression.
Once it is trained, we can use the model to predict the class of new observations:
# Create new observation
new_observation = [[.5, .5, .5, .5]]
# Predict class
model.predict(new_observation)
array([1])
In this example, our observation was predicted to be class 1. Additionally, we can see
the probability that an observation is a member of each class:
# View predicted probabilities
model.predict_proba(new_observation)
array([[ 0.18823041,
0.81176959]])
Our observation had an 18.8% chance of being class 0 and 81.1% chance of being
class 1.
260
|
Chapter 16: Logistic Regression
16.2 Training a Multiclass Classifier
Problem
Given more than two classes, you need to train a classifier model.
Solution
Train a logistic regression in scikit-learn with LogisticRegression using one-vs-rest
or multinomial methods:
# Load libraries
from sklearn.linear_model import LogisticRegression
from sklearn import datasets
from sklearn.preprocessing import StandardScaler
# Load data
iris = datasets.load_iris()
features = iris.data
target = iris.target
# Standardize features
scaler = StandardScaler()
features_standardized = scaler.fit_transform(features)
# Create one-vs-rest logistic regression object
logistic_regression = LogisticRegression(random_state=0, multi_class="ovr")
# Train model
model = logistic_regression.fit(features_standardized, target)
Discussion
On their own, logistic regressions are only binary classifiers, meaning they cannot
handle target vectors with more than two classes. However, two clever extensions to
logistic regression do just that. First, in one-vs-rest logistic regression (OVR) a sepa‐
rate model is trained for each class predicted whether an observation is that class or
not (thus making it a binary classification problem). It assumes that each classifica‐
tion problem (e.g., class 0 or not) is independent.
Alternatively, in multinomial logistic regression (MLR) the logistic function we saw
in Recipe 15.1 is replaced with a softmax function:
P yi = k ∣ X =
e
βkxi
∑Kj = 1 e
β jxi
16.2 Training a Multiclass Classifier |
261
where P(yi = k | X) is the probability of the ith observation’s target value, yi, is class k,
and K is the total number of classes. One practical advantage of the MLR is that its
predicted probabilities using the predict_proba method are more reliable (i.e., better
calibrated).
When using LogisticRegression we can select which of the two techniques we
want, with OVR, ovr, being the default argument. We can switch to an MNL by set‐
ting the argument to multinomial.
16.3 Reducing Variance Through Regularization
Problem
You need to reduce the variance of your logistic regression model.
Solution
Tune the regularization strength hyperparameter, C:
# Load libraries
from sklearn.linear_model import LogisticRegressionCV
from sklearn import datasets
from sklearn.preprocessing import StandardScaler
# Load data
iris = datasets.load_iris()
features = iris.data
target = iris.target
# Standardize features
scaler = StandardScaler()
features_standardized = scaler.fit_transform(features)
# Create decision tree classifier object
logistic_regression = LogisticRegressionCV(
penalty='l2', Cs=10, random_state=0, n_jobs=-1)
# Train model
model = logistic_regression.fit(features_standardized, target)
Discussion
Regularization is a method of penalizing complex models to reduce their variance.
Specifically, a penalty term is added to the loss function we are trying to minimize,
typically the L1 and L2 penalties. In the L1 penalty:
262
|
Chapter 16: Logistic Regression
α
p
∑ βj
j=1
where β j is the parameters of the jth of p features being learned and α is a hyperpara‐
meter denoting the regularization strength. With the L2 penalty:
α
p
2
∑ βj
j=1
Higher values of α increase the penalty for larger parameter values (i.e., more com‐
plex models). scikit-learn follows the common method of using C instead of α where
1
C is the inverse of the regularization strength: C = α . To reduce variance while using
logistic regression, we can treat C as a hyperparameter to be tuned to find the value of
C that creates the best model. In scikit-learn we can use the LogisticRegressionCV
class to efficiently tune C. LogisticRegressionCV’s parameter, Cs, can either accept a
range of values for C to search over (if a list of floats is supplied as an argument) or if
supplied an integer, will generate a list of that many candidate values drawn from a
logarithmic scale between –10,000 and 10,000.
Unfortunately, LogisticRegressionCV does not allow us to search over different pen‐
alty terms. To do this we have to use the less efficient model selection techniques dis‐
cussed in Chapter 12.
16.4 Training a Classifier on Very Large Data
Problem
You need to train a simple classifier model on a very large set of data.
Solution
Train a logistic regression in scikit-learn with LogisticRegression using the stochas‐
tic average gradient (SAG) solver:
# Load libraries
from sklearn.linear_model import LogisticRegression
from sklearn import datasets
from sklearn.preprocessing import StandardScaler
# Load data
iris = datasets.load_iris()
features = iris.data
target = iris.target
16.4 Training a Classifier on Very Large Data
|
263
# Standardize features
scaler = StandardScaler()
features_standardized = scaler.fit_transform(features)
# Create logistic regression object
logistic_regression = LogisticRegression(random_state=0, solver="sag")
# Train model
model = logistic_regression.fit(features_standardized, target)
Discussion
scikit-learn’s LogisticRegression offers a number of techniques for training a logis‐
tic regression, called solvers. Most of the time scikit-learn will select the best solver
automatically for us or warn us that we cannot do something with that solver. How‐
ever, there is one particular case we should be aware of.
While an exact explanation is beyond the bounds of this book (for more information
see Mark Schmidt’s slides in “See Also” on page 264), stochastic average gradient
descent allows us to train a model much faster than other solvers when our data is
very large. However, it is also very sensitive to feature scaling, so standardizing our
features is particularly important. We can set our learning algorithm to use this solver
by setting solver='sag'.
See Also
• Minimizing Finite Sums with the Stochastic Average Gradient Algorithm, Mark
Schmidt
16.5 Handling Imbalanced Classes
Problem
You need to train a simple classifier model.
Solution
Train a logistic regression in scikit-learn using LogisticRegression:
# Load libraries
import numpy as np
from sklearn.linear_model import LogisticRegression
from sklearn import datasets
from sklearn.preprocessing import StandardScaler
# Load data
264
|
Chapter 16: Logistic Regression
iris = datasets.load_iris()
features = iris.data
target = iris.target
# Make class highly imbalanced by removing first 40 observations
features = features[40:,:]
target = target[40:]
# Create target vector indicating if class 0, otherwise 1
target = np.where((target == 0), 0, 1)
# Standardize features
scaler = StandardScaler()
features_standardized = scaler.fit_transform(features)
# Create decision tree classifier object
logistic_regression = LogisticRegression(random_state=0, class_weight="balanced")
# Train model
model = logistic_regression.fit(features_standardized, target)
Discussion
Like many other learning algorithms in scikit-learn, LogisticRegression comes with
a built-in method of handling imbalanced classes. If we have highly imbalanced
classes and have not addressed it during preprocessing, we have the option of using
the class_weight parameter to weight the classes to make certain we have a balanced
mix of each class. Specifically, the balanced argument will automatically weigh classes
inversely proportional to their frequency:
wj =
n
kn j
where wj is the weight to class j, n is the number of observations, nj is the number of
observations in class j, and k is the total number of classes.
16.5 Handling Imbalanced Classes
|
265
CHAPTER 17
Support Vector Machines
17.0 Introduction
To understand support vector machines, we must understand hyperplanes. Formally,
a hyperplane is an n – 1 subspace in an n-dimensional space. While that sounds com‐
plex, it actually is pretty simple. For example, if we wanted to divide a twodimensional space, we’d use a one-dimensional hyperplane (i.e., a line). If we wanted
to divide a three-dimensional space, we’d use a two-dimensional hyperplane (i.e., a
flat piece of paper or a bed sheet). A hyperplane is simply a generalization of that con‐
cept into n dimensions.
Support vector machines classify data by finding the hyperplane that maximizes the
margin between the classes in the training data. In a two-dimensional example with
two classes, we can think of a hyperplane as the widest straight “band” (i.e., line with
margins) that separates the two classes.
In this chapter, we cover training support vector machines in a variety of situations
and dive under the hood to look at how we can extend the approach to tackle com‐
mon problems.
17.1 Training a Linear Classifier
Problem
You need to train a model to classify observations.
Solution
Use a support vector classifier (SVC) to find the hyperplane that maximizes the mar‐
gins between the classes:
267
# Load libraries
from sklearn.svm import LinearSVC
from sklearn import datasets
from sklearn.preprocessing import StandardScaler
import numpy as np
# Load data with only two classes and two features
iris = datasets.load_iris()
features = iris.data[:100,:2]
target = iris.target[:100]
# Standardize features
scaler = StandardScaler()
features_standardized = scaler.fit_transform(features)
# Create support vector classifier
svc = LinearSVC(C=1.0)
# Train model
model = svc.fit(features_standardized, target)
Discussion
scikit-learn’s LinearSVC implements a simple SVC. To get an intuition behind what an
SVC is doing, let us plot out the data and hyperplane. While SVCs work well in high
dimensions, in our solution we only loaded two features and took a subset of obser‐
vations so that the data contains only two classes. This will let us visualize the model.
Recall that SVC attempts to find the hyperplane—a line when we only have two
dimensions—with the maximum margin between the classes. In the following code
we plot the two classes on a two-dimensional space, then draw the hyperplane:
# Load library
from matplotlib import pyplot as plt
# Plot data points and color using their class
color = ["black" if c == 0 else "lightgrey" for c in target]
plt.scatter(features_standardized[:,0], features_standardized[:,1], c=color)
# Create the hyperplane
w = svc.coef_[0]
a = -w[0] / w[1]
xx = np.linspace(-2.5, 2.5)
yy = a * xx - (svc.intercept_[0]) / w[1]
# Plot the hyperplane
plt.plot(xx, yy)
plt.axis("off"), plt.show();
268
|
Chapter 17: Support Vector Machines
In this visualization, all observations of class 0 are black and observations of class 1
are light gray. The hyperplane is the decision boundary deciding how new observa‐
tions are classified. Specifically, any observation above the line will by classified as
class 0 while any observation below the line will be classified as class 1. We can prove
this by creating a new observation in the top-left corner of our visualization, meaning
it should be predicted to be class 0:
# Create new observation
new_observation = [[ -2,
3]]
# Predict class of new observation
svc.predict(new_observation)
array([0])
There are a few things to note about SVCs. First, for the sake of visualization we limi‐
ted our example to a binary example (e.g., only two classes); however, SVCs can work
well with multiple classes. Second, as our visualization shows, the hyperplane is by
definition linear (i.e., not curved). This was okay in this example because the data was
linearly separable, meaning there was a hyperplane that could perfectly separate the
two classes. Unfortunately, in the real world this will rarely be the case.
More typically, we will not be able to perfectly separate classes. In these situations
there is a balance between SVC maximizing the margin of the hyperplane and mini‐
mizing the misclassification. In SVC, the latter is controlled with the hyperparameter
17.1 Training a Linear Classifier |
269
C, the penalty imposed on errors. C is a parameter of the SVC learner and is the pen‐
alty for misclassifying a data point. When C is small, the classifier is okay with mis‐
classified data points (high bias but low variance). When C is large, the classifier is
heavily penalized for misclassified data and therefore bends over backwards to avoid
any misclassified data points (low bias but high variance).
In scikit-learn, C is determined by the parameter C and defaults to C=1.0. We should
treat C has a hyperparameter of our learning algorithm, which we tune using model
selection techniques in Chapter 12.
17.2 Handling Linearly Inseparable Classes Using Kernels
Problem
You need to train a support vector classifier, but your classes are linearly inseparable.
Solution
Train an extension of a support vector machine using kernel functions to create non‐
linear decision boundaries:
# Load libraries
from sklearn.svm import SVC
from sklearn import datasets
from sklearn.preprocessing import StandardScaler
import numpy as np
# Set randomization seed
np.random.seed(0)
# Generate two features
features = np.random.randn(200, 2)
# Use a XOR gate (you don't need to know what this is) to generate
# linearly inseparable classes
target_xor = np.logical_xor(features[:, 0] > 0, features[:, 1] > 0)
target = np.where(target_xor, 0, 1)
# Create a support vector machine with a radial basis function kernel
svc = SVC(kernel="rbf", random_state=0, gamma=1, C=1)
# Train the classifier
model = svc.fit(features, target)
Discussion
A full explanation of support vector machines is outside the scope of this book. How‐
ever, a short explanation is likely beneficial for understanding support vector
270
|
Chapter 17: Support Vector Machines
machines and kernels. For reasons best learned elsewhere, a support vector classifier
can be represented as:
f x = β0 +
∑ αiK xi, xi′
i�S
where β0 is the bias, S is the set of all support vector observations, α are the model
parameters to be learned, and (xi, xi') are pairs of two support vector observations, xi
and xi'. Most importantly, K is a kernel function that compares the similarity between
xi and xi'. Don’t worry if you don’t understand kernel functions. For our purposes, just
realize that K 1) determines the type of hyperplane used to separate our classes and 2)
we create different hyperplanes by using different kernels. For example, if we wanted
the basic linear hyperplane like the one we created in Recipe 17.1, we can use the lin‐
ear kernel:
K xi, xi′ =
p
∑ xi jxi′ j
j=1
where p is the number of features. However, if we wanted a nonlinear decision
boundary, we swap the linear kernel with a polynomial kernel:
K xi, xi′ = 1 +
p
∑ xi jxi′ j
j=1
2
where d is the degree of the polynomial kernel function. Alternatively, we can use one
of the most common kernels in support vectors machines, the radial basis function
kernel:
K xi, xi′ = e
2
− γ∑ pj = 1 xi jx
i′ j
where γ is a hyperparameter and must be greater than zero. The main point of the
preceding explanation is that if we have linearly inseparable data we can swap out a
linear kernel with an alternative kernel to create a nonlinear hyperplane decision
boundary.
We can understand the intuition behind kernels by visualizing a simple example. This
function, based on one by Sebastian Raschka, plots the observations and decision
boundary hyperplane of a two-dimensional space. You do not need to understand
how this function works; I have included it here so you can experiment on your own:
17.2 Handling Linearly Inseparable Classes Using Kernels
|
271
# Plot observations and decision boundary hyperplane
from matplotlib.colors import ListedColormap
import matplotlib.pyplot as plt
def plot_decision_regions(X, y, classifier):
cmap = ListedColormap(("red", "blue"))
xx1, xx2 = np.meshgrid(np.arange(-3, 3, 0.02), np.arange(-3, 3, 0.02))
Z = classifier.predict(np.array([xx1.ravel(), xx2.ravel()]).T)
Z = Z.reshape(xx1.shape)
plt.contourf(xx1, xx2, Z, alpha=0.1, cmap=cmap)
for idx, cl in enumerate(np.unique(y)):
plt.scatter(x=X[y == cl, 0], y=X[y == cl, 1],
alpha=0.8, c=cmap(idx),
marker="+", label=cl)
In our solution, we have data containing two features (i.e., two dimensions) and a tar‐
get vector with the class of each observation. Importantly, the classes are assigned
such that they are linearly inseparable. That is, there is no straight line we can draw
that will divide the two classes. First, let’s create a support vector machine classifier
with a linear kernel:
# Create support vector classifier with a linear kernel
svc_linear = SVC(kernel="linear", random_state=0, C=1)
# Train model
svc_linear.fit(features, target)
SVC(C=1, cache_size=200, class_weight=None, coef0=0.0,
decision_function_shape='ovr', degree=3, gamma='auto', kernel='linear',
max_iter=-1, probability=False, random_state=0, shrinking=True,
tol=0.001, verbose=False)
Next, since we have only two features, we are working in a two-dimensional space
and can visualize the observations, their classes, and our model’s linear hyperplane:
# Plot observations and hyperplane
plot_decision_regions(features, target, classifier=svc_linear)
plt.axis("off"), plt.show();
272
|
Chapter 17: Support Vector Machines
As we can see, our linear hyperplane did very poorly at dividing the two classes! Now,
let’s swap out the linear kernel with a radial basis function kernel and use it to train a
new model:
# Create a support vector machine with a radial basis function kernel
svc = SVC(kernel="rbf", random_state=0, gamma=1, C=1)
# Train the classifier
model = svc.fit(features, target)
And then visualize the observations and hyperplane:
# Plot observations and hyperplane
plot_decision_regions(features, target, classifier=svc)
plt.axis("off"), plt.show();
17.2 Handling Linearly Inseparable Classes Using Kernels
|
273
By using the radial basis function kernel we could create a decision boundary able to
do a much better job of separating the two classes than the linear kernel. This is the
motivation behind using kernels in support vector machines.
In scikit-learn, we can select the kernel we want to use by using the kernel parame‐
ter. Once we select a kernel, we will need to specify the appropriate kernel options
such as the value of d (using the degree parameter) in polynomial kernels and γ
(using the gamma parameter) in radial basis function kernels. We will also need to set
the penalty parameter, C. When training the model, in most cases we should treat all
of these as hyperparameters and use model selection techniques to identify the com‐
bination of their values that produces the model with the best performance.
17.3 Creating Predicted Probabilities
Problem
You need to know the predicted class probabilities for an observation.
Solution
When using scikit-learn’s SVC, set probability=True, train the model, then use pre
dict_proba to see the calibrated probabilities:
274
|
Chapter 17: Support Vector Machines
# Load libraries
from sklearn.svm import SVC
from sklearn import datasets
from sklearn.preprocessing import StandardScaler
import numpy as np
# Load data
iris = datasets.load_iris()
features = iris.data
target = iris.target
# Standardize features
scaler = StandardScaler()
features_standardized = scaler.fit_transform(features)
# Create support vector classifier object
svc = SVC(kernel="linear", probability=True, random_state=0)
# Train classifier
model = svc.fit(features_standardized, target)
# Create new observation
new_observation = [[.4, .4, .4, .4]]
# View predicted probabilities
model.predict_proba(new_observation)
array([[ 0.00588822,
0.96874828,
0.0253635 ]])
Discussion
Many of the supervised learning algorithms we have covered use probability esti‐
mates to predict classes. For example, in k-nearest neighbor, an observation’s k neigh‐
bor’s classes were treated as votes to create a probability that an observation was of
that class. Then the class with the highest probability was predicted. SVC’s use of a
hyperplane to create decision regions does not naturally output a probability estimate
that an observation is a member of a certain class. However, we can in fact output
calibrated class probabilities with a few caveats. In an SVC with two classes, Platt scal‐
ing can be used, wherein first the SVC is trained, and then a separate cross-validated
logistic regression is trained to map the SVC outputs into probabilities:
P y=1∣x =
1
1+e
A* f x + B
where A and B are parameter vectors and f is the ith observation’s signed distance
from the hyperplane. When we have more than two classes, an extension of Platt scal‐
ing is used.
17.3 Creating Predicted Probabilities
|
275
In more practical terms, creating predicted probabilities has two major issues. First,
because we are training a second model with cross-validation, generating predicted
probabilities can significantly increase the time it takes to train our model. Second,
because the predicted probabilities are created using cross-validation, they might not
always match the predicted classes. That is, an observation might be predicted to be
class 1, but have a predicted probability of being class 1 of less than 0.5.
In scikit-learn, the predicted probabilities must be generated when the model is being
trained. We can do this by setting SVC’s probability to True. After the model is
trained, we can output the estimated probabilities for each class using pre
dict_proba.
17.4 Identifying Support Vectors
Problem
You need to identify which observations are the support vectors of the decision
hyperplane.
Solution
Train the model, then use support_vectors_:
# Load libraries
from sklearn.svm import SVC
from sklearn import datasets
from sklearn.preprocessing import StandardScaler
import numpy as np
#Load data with only two classes
iris = datasets.load_iris()
features = iris.data[:100,:]
target = iris.target[:100]
# Standardize features
scaler = StandardScaler()
features_standardized = scaler.fit_transform(features)
# Create support vector classifier object
svc = SVC(kernel="linear", random_state=0)
# Train classifier
model = svc.fit(features_standardized, target)
# View support vectors
model.support_vectors_
array([[-0.5810659 , 0.43490123, -0.80621461, -0.50581312],
[-1.52079513, -1.67626978, -1.08374115, -0.8607697 ],
276
|
Chapter 17: Support Vector Machines
[-0.89430898, -1.46515268,
[-0.5810659 , -1.25403558,
0.30389157,
0.09574666,
0.38157832],
0.55905661]])
Discussion
Support vector machines get their name from the fact that the hyperplane is being
determined by a relatively small number of observations, called the support vectors.
Intuitively, think of the hyperplane as being “carried” by these support vectors. These
support vectors are therefore very important to our model. For example, if we remove
an observation that is not a support vector from the data, the model does not change;
however, if we remove a support vector, the hyperplane will not have the maximum
margin.
After we have trained an SVC, scikit-learn offers us a number of options for identify‐
ing the support vector. In our solution, we used support_vectors_ to output the
actual observations’ features of the four support vectors in our model. Alternatively,
we can view the indices of the support vectors using support_:
model.support_
array([23, 41, 57, 98], dtype=int32)
Finally, we can use n_support_ to find the number of support vectors belonging to
each class:
model.n_support_
array([2, 2], dtype=int32)
17.5 Handling Imbalanced Classes
Problem
You need to train a support vector machine classifier in the presence of imbalanced
classes.
Solution
Increase the penalty for misclassifying the smaller class using class_weight:
# Load libraries
from sklearn.svm import SVC
from sklearn import datasets
from sklearn.preprocessing import StandardScaler
import numpy as np
#Load data with only two classes
iris = datasets.load_iris()
features = iris.data[:100,:]
target = iris.target[:100]
17.5 Handling Imbalanced Classes
|
277
# Make class highly imbalanced by removing first 40 observations
features = features[40:,:]
target = target[40:]
# Create target vector indicating if class 0, otherwise 1
target = np.where((target == 0), 0, 1)
# Standardize features
scaler = StandardScaler()
features_standardized = scaler.fit_transform(features)
# Create support vector classifier
svc = SVC(kernel="linear", class_weight="balanced", C=1.0, random_state=0)
# Train classifier
model = svc.fit(features_standardized, target)
Discussion
In support vector machines, C is a hyperparameter determining the penalty for mis‐
classifying an observation. One method for handling imbalanced classes in support
vector machines is to weight C by classes, so that:
Ck = C * w j
where C is the penalty for misclassification, wj is a weight inversely proportional to
class j’s frequency, and Cj is the C value for class j. The general idea is to increase the
penalty for misclassifying minority classes to prevent them from being “over‐
whelmed” by the majority class.
In scikit-learn, when using SVC we can set the values for Cj automatically by setting
class_weight='balanced'. The balanced argument automatically weighs classes
such that:
wj =
n
kn j
where wj is the weight to class j, n is the number of observations, nj is the number of
observations in class j, and k is the total number of classes.
278
|
Chapter 17: Support Vector Machines
CHAPTER 18
Naive Bayes
18.0 Introduction
Bayes’ theorem is the premier method for understanding the probability of some
event, P(A | B), given some new information, P(B | A), and a prior belief in the proba‐
bility of the event, P(A):
P A∣B =
PB∣A P A
PB
The Bayesian method’s popularity has skyrocketed in the last decade, more and more
rivaling traditional frequentist applications in academia, government, and business.
In machine learning, one application of Bayes’ theorem to classification comes in the
form of the naive Bayes classifier. Naive Bayes classifiers combine a number of desira‐
ble qualities in practical machine learning into a single classifier. These include:
1. An intuitative approach
2. The ability to work with small data
3. Low computation costs for training and prediction
4. Often solid results in a variety of settings.
Specifically, a naive Bayes classifier is based on:
P y ∣ x1, ⋯, x j =
P x1, ⋯x j ∣ y P y
P x1, ⋯, x j
where:
279
• P(y | x1, ⋯, xj) is called the posterior and is the probability that an observation is
class y given the observation’s values for the j features, x1, ⋯, xj.
• P(x1, ...xj | y) is called likelihood and is the likelihood of an observation’s values for
features, x1, ..., xj, given their class, y.
• P(y) is called the prior and is our belief for the probability of class y before look‐
ing at the data.
• P(x1, ..., xj) is called the marginal probability.
In naive Bayes, we compare an observation’s posterior values for each possible class.
Specifically, because the marginal probability is constant across these comparisons,
we compare the numerators of the posterior for each class. For each observation, the
class with the greatest posterior numerator becomes the predicted class, ŷ.
There are two important things to note about naive Bayes classifiers. First, for each
feature in the data, we have to assume the statistical distribution of the likelihood,
P(xj | y). The common distributions are the normal (Gaussian), multinomial, and
Bernoulli distributions. The distribution chosen is often determined by the nature of
features (continuous, binary, etc.). Second, naive Bayes gets its name because we
assume that each feature, and its resulting likelihood, is independent. This “naive”
assumption is frequently wrong, yet in practice does little to prevent building highquality classifiers.
In this chapter we will cover using scikit-learn to train three types of naive Bayes clas‐
sifiers using three different likelihood distributions.
18.1 Training a Classifier for Continuous Features
Problem
You have only continuous features and you want to train a naive Bayes classifier.
Solution
Use a Gaussian naive Bayes classifier in scikit-learn:
# Load libraries
from sklearn import datasets
from sklearn.naive_bayes import GaussianNB
# Load data
iris = datasets.load_iris()
features = iris.data
target = iris.target
# Create Gaussian Naive Bayes object
280
|
Chapter 18: Naive Bayes
classifer = GaussianNB()
# Train model
model = classifer.fit(features, target)
Discussion
The most common type of naive Bayes classifier is the Gaussian naive Bayes. In Gaus‐
sian naive Bayes, we assume that the likelihood of the feature values, x, given an
observation is of class y, follows a normal distribution:
p xj ∣ y =
1
2πσ 2y
−
e
x j − μy
2
2σ 2y
where σy2 and μy are the variance and mean values of feature xj for class y. Because of
the assumption of the normal distribution, Gaussian naive Bayes is best used in cases
when all our features are continuous.
In scikit-learn, we train a Gaussian naive Bayes like any other model using fit, and in
turn can then make predictions about the class of an observation:
# Create new observation
new_observation = [[ 4, 4,
4,
0.4]]
# Predict class
model.predict(new_observation)
array([1])
One of the interesting aspects of naive Bayes classifiers is that they allow us to assign a
prior belief over the respected target classes. We can do this using GaussianNB’s pri
ors parameter, which takes in a list of the probabilities assigned to each class of the
target vector:
# Create Gaussian Naive Bayes object with prior probabilities of each class
clf = GaussianNB(priors=[0.25, 0.25, 0.5])
# Train model
model = classifer.fit(features, target)
If we do not add any argument to the priors parameter, the prior is adjusted based
on the data.
Finally, note that the raw predicted probabilities from Gaussian naive Bayes (output‐
ted using predict_proba) are not calibrated. That is, they should not be believed. If
18.1 Training a Classifier for Continuous Features
|
281
we want to create useful predicted probabilities, we will need to calibrate them using
an isotonic regression or a related method.
See Also
• How the Naive Bayes Classifier Works in Machine Learning, Dataaspirant
18.2 Training a Classifier for Discrete and Count Features
Problem
Given discrete or count data, you need to train a naive Bayes classifier.
Solution
Use a multinomial naive Bayes classifier:
# Load libraries
import numpy as np
from sklearn.naive_bayes import MultinomialNB
from sklearn.feature_extraction.text import CountVectorizer
# Create text
text_data = np.array(['I love Brazil. Brazil!',
'Brazil is best',
'Germany beats both'])
# Create bag of words
count = CountVectorizer()
bag_of_words = count.fit_transform(text_data)
# Create feature matrix
features = bag_of_words.toarray()
# Create target vector
target = np.array([0,0,1])
# Create multinomial naive Bayes object with prior probabilities of each class
classifer = MultinomialNB(class_prior=[0.25, 0.5])
# Train model
model = classifer.fit(features, target)
Discussion
Multinomial naive Bayes works similarly to Gaussian naive Bayes, but the features are
assumed to be multinomially distributed. In practice, this means that this classifier is
commonly used when we have discrete data (e.g., movie ratings ranging from 1 to 5).
282
|
Chapter 18: Naive Bayes
One of the most common uses of multinomial naive Bayes is text classification using
bags of words or tf-idf approaches (see Recipes 6.8 and 6.9).
In our solution, we created a toy text dataset of three observations, and converted the
text strings into a bag-of-words feature matrix and an accompanying target vector.
We then used MultinomialNB to train a model while defining the prior probabilities
for the two classes (pro-brazil and pro-germany).
MultinomialNB works similarly to GaussianNB; models are trained using fit, and
observations can be predicted using predict:
# Create new observation
new_observation = [[0, 0, 0, 1, 0, 1, 0]]
# Predict new observation's class
model.predict(new_observation)
array([0])
If class_prior is not specified, prior probabilities are learned using the data. How‐
ever, if we want a uniform distribution to be used as the prior, we can set
fit_prior=False.
Finally, MultinomialNB contains an additive smoothing hyperparameter, alpha, that
should be tuned. The default value is 1.0, with 0.0 meaning no smoothing takes
place.
18.3 Training a Naive Bayes Classifier for Binary Features
Problem
You have binary feature data and need to train a naive Bayes classifier.
Solution
Use a Bernoulli naive Bayes classifier:
# Load libraries
import numpy as np
from sklearn.naive_bayes import BernoulliNB
# Create three binary features
features = np.random.randint(2, size=(100, 3))
# Create a binary target vector
target = np.random.randint(2, size=(100, 1)).ravel()
# Create Bernoulli Naive Bayes object with prior probabilities of each class
classifer = BernoulliNB(class_prior=[0.25, 0.5])
18.3 Training a Naive Bayes Classifier for Binary Features
|
283
# Train model
model = classifer.fit(features, target)
Discussion
The Bernoulli naive Bayes classifier assumes that all our features are binary such that
they take only two values (e.g., a nominal categorical feature that has been one-hot
encoded). Like its multinomial cousin, Bernoulli naive Bayes is often used in text
classification, when our feature matrix is simply the presence or absence of a word in
a document. Furthermore, like MultinomialNB, BernoulliNB has an additive smooth‐
ing hyperparameter, alpha, we will want to tune using model selection techniques.
Finally, if we want to use priors we can use the class_prior parameter with a list
containing the prior probabilities for each class. If we want to specify a uniform prior,
we can set fit_prior=False:
model_uniform_prior = BernoulliNB(class_prior=None, fit_prior=True)
18.4 Calibrating Predicted Probabilities
Problem
You want to calibrate the predicted probabilities from naive Bayes classifiers so they
are interpretable.
Solution
Use CalibratedClassifierCV:
# Load libraries
from sklearn import datasets
from sklearn.naive_bayes import GaussianNB
from sklearn.calibration import CalibratedClassifierCV
# Load data
iris = datasets.load_iris()
features = iris.data
target = iris.target
# Create Gaussian Naive Bayes object
classifer = GaussianNB()
# Create calibrated cross-validation with sigmoid calibration
classifer_sigmoid = CalibratedClassifierCV(classifer, cv=2, method='sigmoid')
# Calibrate probabilities
classifer_sigmoid.fit(features, target)
# Create new observation
284
|
Chapter 18: Naive Bayes
new_observation = [[ 2.6,
2.6,
2.6,
0.4]]
# View calibrated probabilities
classifer_sigmoid.predict_proba(new_observation)
array([[ 0.31859969,
0.63663466,
0.04476565]])
Discussion
Class probabilities are a common and useful part of machine learning models. In
scikit-learn, most learning algortihms allow us to see the predicted probabilities of
class membership using predict_proba. This can be extremely useful if, for instance,
we want to only predict a certain class if the model predicts the probability that it is
that class is over 90%. However, some models, including naive Bayes classifiers, out‐
put probabilities that are not based on the real world. That is, predict_proba might
predict an observation has a 0.70 chance of being a certain class, when the reality is
that it is 0.10 or 0.99. Specifically in naive Bayes, while the ranking of predicted prob‐
abilities for the different target classes is valid, the raw predicted probabilities tend to
take on extreme values close to 0 and 1.
To obtain meaningful predicted probabilities we need conduct what is called calibra‐
tion. In scikit-learn we can use the CalibratedClassifierCV class to create wellcalibrated predicted probabilities using k-fold cross-validation. In CalibratedClassi
fierCV the training sets are used to train the model and the test set is used to calibrate
the predicted probabilities. The returned predicted probabilities are the average of the
k-folds.
Using our solution we can see the difference between raw and well-calibrated predic‐
ted probabilities. In our solution, we created a Gaussian naive Bayes classifier. If we
train that classifier and then predict the class probabilities for a new observation, we
can see very extreme probability estimates:
# Train a Gaussian naive Bayes then predict class probabilities
classifer.fit(features, target).predict_proba(new_observation)
array([[
2.58229098e-04,
9.99741447e-01,
3.23523643e-07]])
However, if after we calibrate the predicted probabilities (which we did in our solu‐
tion), we get very different results:
# View calibrated probabilities
classifer_sigmoid.predict_proba(new_observation)
array([[ 0.31859969,
0.63663466,
0.04476565]])
CalibratedClassifierCV offers two calibration methods—Platt’s sigmoid model and
isotonic regression—defined by the method paramenter. While we don’t have the
space to go into the specifics, because isotonic regression is nonparametric it tends to
overfit when sample sizes are very small (e.g., 100 observations). In our solution we
18.4 Calibrating Predicted Probabilities
|
285
used the Iris dataset with 150 observations and therefore used the Platt’s sigmoid
model.
286
|
Chapter 18: Naive Bayes
CHAPTER 19
Clustering
19.0 Introduction
In much of this book we have looked at supervised machine learning—where we have
access to both the features and the target. This is, unfortunately, not always the case.
Frequently, we run into situations where we only know the features. For example,
imagine we have records of sales from a grocery store and we want to break up sales
by whether or not the shopper is a member of a discount club. This would be impos‐
sible using supervised learning because we don’t have a target to train and evaluate
our models. However, there is another option: unsupervised learning. If the behavior
of discount club members and nonmembers in the grocery store is actually disparate,
then the average difference in behavior between two members will be smaller than
the average difference in behavior between a member and nonmember shopper. Put
another way, there will be two clusters of observations.
The goal of clustering algorithms is to identify those latent groupings of observations,
which if done well, allow us to predict the class of observations even without a target
vector. There are many clustering algorithms and they have a wide variety of
approaches to identifying the clusters in data. In this chapter, we will cover a selection
of clustering algorithms using scikit-learn and how to use them in practice.
19.1 Clustering Using K-Means
Problem
You want to group observations into k groups.
287
Solution
Use k-means clustering:
# Load libraries
from sklearn import datasets
from sklearn.preprocessing import StandardScaler
from sklearn.cluster import KMeans
# Load data
iris = datasets.load_iris()
features = iris.data
# Standardize features
scaler = StandardScaler()
features_std = scaler.fit_transform(features)
# Create k-mean object
cluster = KMeans(n_clusters=3, random_state=0, n_jobs=-1)
# Train model
model = cluster.fit(features_std)
Discussion
k-means clustering is one of the most common clustering techniques. In k-means
clustering, the algorithm attempts to group observations into k groups, with each
group having roughly equal variance. The number of groups, k, is specified by the
user as a hyperparameter. Specifically, in k-means:
1. k cluster “center” points are created at random locations.
2. For each observation:
a. The distance between each observation and the k center points is calculated.
b. The observation is assigned to the cluster of the nearest center point.
3. The center points are moved to the means (i.e., centers) of their respective
clusters.
4. Steps 2 and 3 are repeated until no observation changes in cluster membership.
At this point the algorithm is considered converged and stops.
It is important to note three things about k-means. First, k-means clustering assumes
the clusters are convex shaped (e.g., a circle, a sphere). Second, all features are equally
scaled. In our solution, we standardized the features to meet this assumption. Third,
the groups are balanced (i.e., have roughly the same number of observations). If we
suspect that we cannot meet these assumptions, we might try other clustering
approaches.
288
|
Chapter 19: Clustering
In scikit-learn, k-means clustering is implemented in the KMeans class. The most
important parameter is n_clusters, which sets the number of clusters k. In some sit‐
uations, the nature of the data will determine the value for k (e.g., data on a school’s
students will have one cluster per grade), but often we don’t know the number of
clusters. In these cases, we will want to select k based on using some criteria. For
example, silhouette coefficients (see Recipe 11.9) measure the similarity within clus‐
ters compared with the similarity between clusters. Furthermore, because k-means
clustering is computationally expensive, we might want to take advantage of all the
cores on our computer. We can do this by setting n_jobs=-1.
In our solution, we cheated a little and used the Iris flower data, in which we know
there are three classes. Therefore, we set k = 3. We can use labels_ to see the predic‐
ted classes of each observation:
# View predict class
model.labels_
array([1,
1,
1,
2,
2,
0,
2,
1,
1,
1,
0,
2,
0,
0,
1,
1,
1,
2,
2,
0,
0,
1,
1,
1,
2,
2,
0,
0,
1,
1,
0,
2,
2,
2,
2,
1,
1,
0,
2,
2,
0,
0,
1,
1,
0,
0,
2,
2,
0,
1,
1,
2,
0,
2,
0,
0,
1,
1,
2,
0,
0,
2,
2,
1,
1,
2,
2,
2,
0,
0,
1,
1,
0,
2,
0,
0,
0,
1, 1, 1, 1, 1, 1,
1, 1, 1, 1, 1, 1,
2, 2, 2, 2, 2, 2,
2, 2, 2, 2, 2, 0,
0, 0, 0, 2, 0, 0,
2, 0, 0, 0, 0, 0,
2], dtype=int32)
1,
1,
2,
0,
0,
0,
1,
1,
2,
2,
0,
2,
1,
1,
0,
2,
0,
2,
1,
1,
2,
2,
0,
0,
1,
1,
2,
2,
2,
0,
1,
1,
2,
2,
2,
0,
If we compare this to the observation’s true class we can see that despite the difference
in class labels (i.e., 1, 2, and 3), k-means did reasonably well:
# View true class
iris.target
array([0,
0,
0,
1,
1,
2,
2,
0,
0,
0,
1,
1,
2,
2,
0,
0,
0,
1,
1,
2,
2,
0,
0,
0,
1,
1,
2,
2,
0,
0,
1,
1,
1,
2,
2,
0,
0,
1,
1,
1,
2,
2,
0,
0,
1,
1,
1,
2,
2,
0,
0,
1,
1,
1,
2,
2,
0,
0,
1,
1,
2,
2,
2,
0,
0,
1,
1,
2,
2,
2,
0,
0,
1,
1,
2,
2,
2,
0, 0,
0, 0,
1, 1,
1, 1,
2, 2,
2, 2,
2])
0,
0,
1,
1,
2,
2,
0,
0,
1,
1,
2,
2,
0,
0,
1,
1,
2,
2,
0,
0,
1,
1,
2,
2,
0,
0,
1,
1,
2,
2,
0,
0,
1,
1,
2,
2,
0,
0,
1,
1,
2,
2,
0,
0,
1,
1,
2,
2,
0,
0,
1,
1,
2,
2,
0,
0,
1,
1,
2,
2,
However, as you might imagine, the performance of k-means drops considerably,
even critically, if we select the wrong number of clusters.
Finally, as with other scikit-learns we can use the trained cluster to predict the value
of new observations:
# Create new observation
new_observation = [[0.8, 0.8, 0.8, 0.8]]
# Predict observation's cluster
model.predict(new_observation)
array([0], dtype=int32)
19.1 Clustering Using K-Means
|
289
The observation is predicted to belong to the cluster whose center point is closest. We
can even use cluster_centers_ to see those center points:
# View cluster centers
model.cluster_centers_
array([[ 1.13597027, 0.09659843, 0.996271 , 1.01717187],
[-1.01457897, 0.84230679, -1.30487835, -1.25512862],
[-0.05021989, -0.88029181, 0.34753171, 0.28206327]])
See Also
• Introduction to K-means Clustering, DataScience.com
19.2 Speeding Up K-Means Clustering
Problem
You want to group observations into k groups, but k-means takes too long.
Solution
Use mini-batch k-means:
# Load libraries
from sklearn import datasets
from sklearn.preprocessing import StandardScaler
from sklearn.cluster import MiniBatchKMeans
# Load data
iris = datasets.load_iris()
features = iris.data
# Standardize features
scaler = StandardScaler()
features_std = scaler.fit_transform(features)
# Create k-mean object
cluster = MiniBatchKMeans(n_clusters=3, random_state=0, batch_size=100)
# Train model
model = cluster.fit(features_std)
Discussion
Mini-batch k-means works similarly to the k-means algorithm discussed in Recipe
19.1. Without going into too much detail, the difference is that in mini-batch
k-means the most computationally costly step is conducted on only a random sample
290
|
Chapter 19: Clustering
of observations as opposed to all observations. This approach can significantly reduce
the time required for the algorithm to find convergence (i.e., fit the data) with only a
small cost in quality.
MiniBatchKMeans works similarly to KMeans, with one significant difference: the
batch_size parameter. batch_size controls the number of randomly selected obser‐
vations in each batch. The larger the size of the batch, the more computationally
costly the training process.
19.3 Clustering Using Meanshift
Problem
You want to group observations without assuming the number of clusters or their
shape.
Solution
Use meanshift clustering:
# Load libraries
from sklearn import datasets
from sklearn.preprocessing import StandardScaler
from sklearn.cluster import MeanShift
# Load data
iris = datasets.load_iris()
features = iris.data
# Standardize features
scaler = StandardScaler()
features_std = scaler.fit_transform(features)
# Create meanshift object
cluster = MeanShift(n_jobs=-1)
# Train model
model = cluster.fit(features_std)
Discussion
One of the disadvantages of k-means clustering we discussed previously is that we
needed to set the number of clusters, k, prior to training, and the method made
assumptions about the shape of the clusters. One clustering algorithm without these
limitations is meanshift.
Meanshift is a simple concept, but somewhat difficult to explain. Therefore, an anal‐
ogy might be the best approach. Imagine a very foggy football field (i.e., a two19.3 Clustering Using Meanshift
|
291
dimensional feature space) with 100 people standing on it (i.e., our observations).
Because it is foggy, a person can only see a short distance. Every minute each person
looks around and takes a step in the direction of the most people they can see. As
time goes on, people start to group up as they repeatedly take steps toward larger and
larger crowds. The end result is clusters of people around the field. People are
assigned to the clusters in which they end up.
scikit-learn’s actual implementation of meanshift, MeanShift, is more complex but
follows the same basic logic. MeanShift has two important parameters we should be
aware of. First, bandwidth sets the radius of the area (i.e., kernel) an observation uses
to determine the direction to shift. In our analogy, bandwidth was how far a person
could see through the fog. We can set this parameter manually, but by default a rea‐
sonable bandwidth is estimated automatically (with a significant increase in computa‐
tional cost). Second, sometimes in meanshift there are no other observations within
an observation’s kernel. That is, a person on our football field cannot see a single
other person. By default, MeanShift assigns all these “orphan” observations to the
kernel of the nearest observation. However, if we want to leave out these orphans, we
can set cluster_all=False wherein orphan observations are given the label of -1.
See Also
• The mean shift clustering algorithm, EFAVDB
19.4 Clustering Using DBSCAN
Problem
You want to group observations into clusters of high density.
Solution
Use DBSCAN clustering:
# Load libraries
from sklearn import datasets
from sklearn.preprocessing import StandardScaler
from sklearn.cluster import DBSCAN
# Load data
iris = datasets.load_iris()
features = iris.data
# Standardize features
scaler = StandardScaler()
features_std = scaler.fit_transform(features)
292
|
Chapter 19: Clustering
# Create meanshift object
cluster = DBSCAN(n_jobs=-1)
# Train model
model = cluster.fit(features_std)
Discussion
DBSCAN is motivated by the idea that clusters will be areas where many observations
are densely packed together and makes no assumptions of cluster shape. Specifically,
in DBSCAN:
1. A random observation, xi, is chosen.
2. If xi has a minimum number of close neighbors, we consider it to be part of a
cluster.
3. Step 2 is repeated recursively for all of xi’s neighbors, then neighbor’s neighbor,
and so on. These are the cluster’s core observations.
4. Once step 3 runs out of nearby observations, a new random point is chosen (i.e.,
restarting step 1).
Once this is complete, we have a set of core observations for a number of clusters.
Finally, any observation close to a cluster but not a core sample is considered part of a
cluster, while any observation not close to the cluster is labeled an outlier.
DBSCAN has three main parameters to set:
eps
The maximum distance from an observation for another observation to be con‐
sidered its neighbor.
min_samples
The minimum number of observations less than eps distance from an observa‐
tion for it to be considered a core observation.
metric
The distance metric used by eps—for example, minkowski or euclidean (note
that if Minkowski distance is used, the parameter p can be used to set the power
of the Minkowski metric).
If we look at the clusters in our training data we can see two clusters have been identi‐
fied, 0 and 1, while outlier observations are labeled -1:
# Show cluster membership
model.labels_
19.4 Clustering Using DBSCAN
|
293
array([ 0,
0,
0,
1,
-1,
-1,
1,
-1,
-1,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, -1,
0, 0, 0, 0, -1, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, -1, 0, 0, 0, 0, 0, 0, 0,
1, 1, 1, 1, -1, -1, 1, -1, -1, 1, -1, 1, 1, 1,
1, 1, 1, -1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
1, -1, 1, 1, 1, 1, 1, -1, 1, 1, 1, 1, -1, 1,
1, 1, -1, -1, -1, -1, -1, 1, 1, 1, 1, -1, 1, 1,
1, 1, -1, 1, 1, -1, 1, 1, 1, -1, -1, -1, 1, 1,
1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, -1, 1])
-1, 0,
-1, -1,
0, 1,
1, 1,
1, 1,
-1, 1,
-1, -1,
1, -1,
See Also
• DBSCAN, Wikipedia
19.5 Clustering Using Hierarchical Merging
Problem
You want to group observations using a hierarchy of clusters.
Solution
Use agglomerative clustering:
# Load libraries
from sklearn import datasets
from sklearn.preprocessing import StandardScaler
from sklearn.cluster import AgglomerativeClustering
# Load data
iris = datasets.load_iris()
features = iris.data
# Standardize features
scaler = StandardScaler()
features_std = scaler.fit_transform(features)
# Create meanshift object
cluster = AgglomerativeClustering(n_clusters=3)
# Train model
model = cluster.fit(features_std)
Discussion
Agglomerative clustering is a powerful, flexible hierarchical clustering algorithm. In
agglomerative clustering, all observations start as their own clusters. Next, clusters
meeting some criteria are merged together. This process is repeated, growing clusters
294
|
Chapter 19: Clustering
until some end point is reached. In scikit-learn, AgglomerativeClustering uses the
linkage parameter to determine the merging strategy to minimize the following:
1. Variance of merged clusters (ward)
2. Average distance between observations from pairs of clusters (average)
3. Maximum distance between observations from pairs of clusters (complete)
Two other parameters are useful to know. First, the affinity parameter determines
the distance metric used for linkage (minkowski, euclidean, etc.). Second, n_clus
ters sets the number of clusters the clustering algorithm will attempt to find. That is,
clusters are successively merged until there are only n_clusters remaining.
As with other clustering algorithms we have covered, we can use labels_ to see the
cluster in which every observation is assigned:
# Show cluster membership
model.labels_
array([1,
1,
1,
2,
2,
0,
0,
1,
1,
1,
0,
2,
0,
0,
1,
1,
1,
0,
2,
0,
0,
1,
1,
1,
0,
2,
0,
0,
1,
1,
0,
0,
2,
2,
0,
1,
1,
0,
0,
0,
0,
0,
1,
1,
0,
0,
2,
0,
0,
1,
1,
2,
0,
2,
0,
0,
1,
1,
0,
0,
0,
0,
0,
1,
1,
2,
0,
0,
0,
0,
1,
1,
0,
2,
0,
0,
0,
1, 1,
1, 1,
2, 0,
2, 2,
0, 0,
0, 0,
0])
1,
1,
2,
2,
0,
0,
1,
1,
2,
0,
2,
0,
1,
1,
0,
2,
0,
0,
1,
1,
2,
0,
0,
0,
1,
1,
0,
0,
0,
0,
1,
2,
2,
2,
0,
0,
1,
1,
0,
2,
0,
0,
1,
1,
2,
2,
0,
0,
1,
1,
2,
2,
0,
0,
1,
1,
2,
0,
0,
0,
19.5 Clustering Using Hierarchical Merging
|
295
CHAPTER 20
Neural Networks
20.0 Introduction
At the heart of neural networks is the unit (also called a node or neuron). A unit takes
in one or more inputs, multiplies each input by a parameter (also called a weight),
sums the weighted input’s values along with some bias value (typically 1), and then
feeds the value into an activation function. This output is then sent forward to the
other neurals deeper in the neural network (if they exist).
Feedforward neural networks—also called multilayer perceptron—are the simplest
artificial neural network used in any real-world setting. Neural networks can be
visualized as a series of connected layers that form a network connecting an observa‐
tion’s feature values at one end, and the target value (e.g., observation’s class) at the
other end. The name feedforward comes from the fact that an observation’s feature
values are fed “forward” through the network, with each layer successively transform‐
ing the feature values with the goal that the output at the end is the same as the tar‐
get’s value.
Specifically, feedforward neural networks contain three types of layers of units. At the
start of the neural network is an input layer where each unit contains an observation’s
value for a single feature. For example, if an observation has 100 features, the input
layer has 100 nodes. At the end of the neural network is the output layer, which trans‐
forms the output of the hidden layers into values useful for the task at hand. For
example, if our goal was binary classification, we could use an output layer with a sin‐
gle unit that uses a sigmoid function to scale its own output to between 0 and 1, rep‐
resenting a predicted class probability. Between the input and output layers are the
so-called “hidden” layers (which aren’t hidden at all). These hidden layers successively
transform the feature values from the input layer to something that, once processed
by the output layer, resembles the target class. Neural networks with many hidden
297
layers (e.g., 10, 100, 1,000) are considered “deep” networks and their application is
called deep learning.
Neural networks are typically created with all parameters initialized as small random
values from a Gaussian or normal uniform. Once an observation (or more often a set
number of observations called a batch) is fed through the network, the outputted
value is compared with the observation’s true value using a loss function. This is
called forward propagation. Next an algorithm goes “backward” through the network
identifying how much each parameter contributed to the error between the predicted
and true values, a process called backpropagation. At each parameter, the optimiza‐
tion algorithm determines how much each weight should be adjusted to improve the
output.
Neural networks learn by repeating this process of forward propagation and backpro‐
pagation for every observation in the training data multiple times (each time all
observations have been sent through the network is called an epoch and training typi‐
cally consists of multiple epochs), iteratively updating the values of the parameters.
In this chapter, we will use the popular Python library Keras to build, train, and eval‐
uate a variety of neural networks. Keras is a high-level library, using other libraries
like TensorFlow and Theano as its “engine.” For us the advantage of Keras is that we
can focus on network design and training, leaving the specifics of the tensor opera‐
tions to the other libraries.
Neural networks created using Keras code can be trained using both CPUs (i.e., on
your laptop) and GPUs (i.e., on a specialized deep learning computer). In the real
world with real data, it is highly advisable to train neural networks using GPUs; how‐
ever, for the sake of learning, all the neural networks in this book are small and sim‐
ple enough to be trained on your laptop in only a few minutes. Just be aware that
when we have larger networks and more training data, training using CPUs is signifi‐
cantly slower than training using GPUs.
20.1 Preprocessing Data for Neural Networks
Problem
You want to preprocess data for use in a neural network.
Solution
Standardize each feature using scikit-learn’s StandardScaler:
# Load libraries
from sklearn import preprocessing
import numpy as np
298
|
Chapter 20: Neural Networks
# Create feature
features = np.array([[-100.1,
[-200.2,
[5000.5,
[6000.6,
[9000.9,
3240.1],
-234.1],
150.1],
-125.1],
-673.1]])
# Create scaler
scaler = preprocessing.StandardScaler()
# Transform the feature
features_standardized = scaler.fit_transform(features)
# Show feature
features_standardized
array([[-1.12541308,
[-1.15329466,
[ 0.29529406,
[ 0.57385917,
[ 1.40955451,
1.96429418],
-0.50068741],
-0.22809346],
-0.42335076],
-0.81216255]])
Discussion
While this recipe is very similar to Recipe 4.3, it is worth repeating because of how
important it is for neural networks. Typically, a neural network’s parameters are ini‐
tialized (i.e., created) as small random numbers. Neural networks often behave poorly
when the feature values are much larger than parameter values. Furthermore, since
an observation’s feature values are combined as they pass through individual units, it
is important that all features have the same scale.
For these reasons, it is best practice (although not always necessary; for example,
when we have all binary features) to standardize each feature such that the feature’s
values have the mean of 0 and the standard deviation of 1. This can be easily accom‐
plished with scikit-learn’s StandardScaler.
You can see the effect of the standardization by checking the mean and standard devi‐
ation of our first features:
# Print mean and standard deviation
print("Mean:", round(features_standardized[:,0].mean()))
print('"Standard deviation:", features_standardized[:,0].std())
Mean: 0.0
Standard deviation: 1.0
20.1 Preprocessing Data for Neural Networks
|
299
20.2 Designing a Neural Network
Problem
You want to design a neural network.
Solution
Use Keras’ Sequential model:
# Load libraries
from keras import models
from keras import layers
# Start neural network
network = models.Sequential()
# Add fully connected layer with a ReLU activation function
network.add(layers.Dense(units=16, activation="relu", input_shape=(10,)))
# Add fully connected layer with a ReLU activation function
network.add(layers.Dense(units=16, activation="relu"))
# Add fully connected layer with a sigmoid activation function
network.add(layers.Dense(units=1, activation="sigmoid"))
# Compile neural network
network.compile(loss="binary_crossentropy", # Cross-entropy
optimizer="rmsprop", # Root Mean Square Propagation
metrics=["accuracy"]) # Accuracy performance metric
Using TensorFlow backend.
Discussion
Neural networks consist of layers of units. However, there is incredible variety in the
types of layers and how they are combined to form the network’s architecture. At
present, while there are commonly used architecture patterns (which we will cover in
this chapter), the truth is that selecting the right architecture is mostly an art and the
topic of much research.
To construct a feedforward neural network in Keras, we need to make a number of
choices about both the network architecture and training process. Remember that
each unit in the hidden layers:
1. Receives a number of inputs.
2. Weights each input by a parameter value.
3. Sums together all weighted inputs along with some bias (typically 1).
300
|
Chapter 20: Neural Networks
4. Most often then applies some function (called an activation function).
5. Sends the output on to units in the next layer.
First, for each layer in the hidden and output layers we must define the number of
units to include in the layer and the activation function. Overall, the more units we
have in a layer, the more our network is able to learn complex patterns. However,
more units might make our network overfit the training data in a way detrimental to
the performance on the test data.
For hidden layers, a popular activation function is the rectified linear unit (ReLU):
f z = max 0, z
where z is the sum of the weighted inputs and bias. As we can see, if z is greater than
0, the activation function returns z; otherwise, the function returns 0. This simple
activation function has a number of desirable properties (a discussion of which is
beyond the scope of this book) and this has made it a popular choice in neural net‐
works. We should be aware, however, that many dozens of activation functions exist.
Second, we need to define the number of hidden layers to use in the network. More
layers allow the network to learn more complex relationships, but with a computa‐
tional cost.
Third, we have to define the structure of the activation function (if any) of the output
layer. The nature of the output function is often determined by the goal of the net‐
work. Here are some common output layer patterns:
Binary classification
One unit with a sigmoid activation function.
Multiclass classification
k units (where k is the number of target classes) and a softmax activation func‐
tion.
Regression
One unit with no activation function.
Fourth, we need to define a loss function (the function that measures how well a pre‐
dicted value matches the true value); this is again often determined by the problem
type:
Binary classification
Binary cross-entropy.
Multiclass classification
Categorical cross-entropy.
20.2 Designing a Neural Network
|
301
Regression
Mean square error.
Fifth, we have to define an optimizer, which intuitively can be thought of as our strat‐
egy “walking around” the loss function to find the parameter values that produce the
lowest error. Common choices for optimizers are stochastic gradient descent, sto‐
chastic gradient descent with momentum, root mean square propagation, and adap‐
tive moment estimation (more information on these optimizers in “See Also” on page
302).
Sixth, we can select one or more metrics to use to evaluate the performance, such as
accuracy.
Keras offers two ways for creating neural networks. Keras’ sequential model creates
neural networks by stacking together layers. An alternative method for creating neu‐
ral networks is called the functional API, but that is more for researchers rather than
practitioners.
In our solution, we created a two-layer neural network (when counting layers we
don’t include the input layer because it does not have any parameters to learn) using
Keras’ sequential model. Each layer is “dense” (also called fully connected), meaning
that all the units in the previous layer are connected to all the neurals in the next
layer. In the first hidden layer we set units=16, meaning that layer contains 16 units
with ReLU activation functions: activation='relu'. In Keras, the first hidden layer
of any network has to include an input_shape parameter, which is the shape of fea‐
ture data. For example, (10,) tells the first layer to expect each observation to have 10
feature values. Our second layer is the same as the first, without the need for the
input_shape parameter. This network is designed for binary classification so the out‐
put layer contains only one unit with a sigmoid activation function, which constrains
the output to between 0 and 1 (representing the probability an observation is class 1).
Finally, before we can train our model, we need to tell Keras how we want our net‐
work to learn. We do this using the compile method, with our optimization algo‐
rithm (RMSProp), loss function (binary_crossentropy), and one or more perfor‐
mance metrics.
See Also
• Losses, Keras
• Loss functions for classification, Wikipedia
• On Loss Functions for Deep Neural Networks in Classification, Katarzyna Jano‐
cha, Wojciech Marian Czarnecki
302
|
Chapter 20: Neural Networks
20.3 Training a Binary Classifier
Problem
You want to train a binary classifier neural network.
Solution
Use Keras to construct a feedforward neural network and train it using the fit
method:
# Load libraries
import numpy as np
from keras.datasets import imdb
from keras.preprocessing.text import Tokenizer
from keras import models
from keras import layers
# Set random seed
np.random.seed(0)
# Set the number of features we want
number_of_features = 1000
# Load data and target vector from movie review data
(data_train, target_train), (data_test, target_test) = imdb.load_data(
num_words=number_of_features)
# Convert movie review data to one-hot encoded feature matrix
tokenizer = Tokenizer(num_words=number_of_features)
features_train = tokenizer.sequences_to_matrix(data_train, mode="binary")
features_test = tokenizer.sequences_to_matrix(data_test, mode="binary")
# Start neural network
network = models.Sequential()
# Add fully connected layer with a ReLU activation function
network.add(layers.Dense(units=16, activation="relu", input_shape=(
number_of_features,)))
# Add fully connected layer with a ReLU activation function
network.add(layers.Dense(units=16, activation="relu"))
# Add fully connected layer with a sigmoid activation function
network.add(layers.Dense(units=1, activation="sigmoid"))
# Compile neural network
network.compile(loss="binary_crossentropy", # Cross-entropy
optimizer="rmsprop", # Root Mean Square Propagation
metrics=["accuracy"]) # Accuracy performance metric
20.3 Training a Binary Classifier
|
303
# Train neural network
history = network.fit(features_train, # Features
target_train, # Target vector
epochs=3, # Number of epochs
verbose=1, # Print description after each epoch
batch_size=100, # Number of observations per batch
validation_data=(features_test, target_test)) # Test data
Using TensorFlow backend.
Train on 25000 samples, validate on 25000 samples
Epoch 1/3
25000/25000 [==============================]
- 1s - loss: 0.4215 - acc: 0.8105 - val_loss: 0.3386 - val_acc: 0.8555
Epoch 2/3
25000/25000 [==============================]
- 1s - loss: 0.3242 - acc: 0.8645 - val_loss: 0.3258 - val_acc: 0.8633
Epoch 3/3
25000/25000 [==============================]
- 1s - loss: 0.3121 - acc: 0.8699 - val_loss: 0.3265 - val_acc: 0.8599
Discussion
In Recipe 20.2, we discussed how to construct a neural network using Keras’ sequen‐
tial model. In this recipe we train that neural network using real data. Specifically, we
use 50,000 movie reviews (25,000 as training data, 25,000 held out for testing), cate‐
gorized as positive or negative. We convert the text of the reviews into 5,000 binary
features indicating the presence of one of the 1,000 most frequent words. Put more
simply, our neural networks will use 25,000 observations, each with 1,000 features, to
predict if a movie review is positive or negative.
The neural network we are using is the same as the one in Recipe 20.2 (see there for a
detailed explanation). The only addition is that in that recipe we only created the neu‐
ral network, we didn’t train it.
In Keras, we train our neural network using the fit method. There are six significant
parameters to define. The first two parameters are the features and target vector of
the training data. We can view the shape of feature matrix using shape:
# View shape of feature matrix
features_train.shape
(25000, 1000)
The epochs parameter defines how many epochs to use when training the data. ver
bose determines how much information is outputted during the training process,
with 0 being no output, 1 outputting a progress bar, and 2 one log line per epoch.
batch_size sets the number of observations to propagate through the network before
updating the parameters.
304
|
Chapter 20: Neural Networks
Finally, we held out a test set of data to use to evaluate the model. These test features
and test target vector can be arguments of validation_data, which will use them for
evaluation. Alternatively, we could have used validation_split to define what frac‐
tion of the training data we want to hold out for evaluation.
In scikit-learn the fit method returned a trained model, but in Keras the fit method
returns a History object containing the loss values and performance metrics at each
epoch.
20.4 Training a Multiclass Classifier
Problem
You want to train a multiclass classifier neural network.
Solution
Use Keras to construct a feedforward neural network with an output layer with soft‐
max activation functions:
# Load libraries
import numpy as np
from keras.datasets import reuters
from keras.utils.np_utils import to_categorical
from keras.preprocessing.text import Tokenizer
from keras import models
from keras import layers
# Set random seed
np.random.seed(0)
# Set the number of features we want
number_of_features = 5000
# Load feature and target data
data = reuters.load_data(num_words=number_of_features)
(data_train, target_vector_train), (data_test, target_vector_test) = data
# Convert feature data to a one-hot encoded feature matrix
tokenizer = Tokenizer(num_words=number_of_features)
features_train = tokenizer.sequences_to_matrix(data_train, mode="binary")
features_test = tokenizer.sequences_to_matrix(data_test, mode="binary")
# One-hot encode target vector to create a target matrix
target_train = to_categorical(target_vector_train)
target_test = to_categorical(target_vector_test)
# Start neural network
network = models.Sequential()
20.4 Training a Multiclass Classifier
|
305
# Add fully connected layer with a ReLU activation function
network.add(layers.Dense(units=100,
activation="relu",
input_shape=(number_of_features,)))
# Add fully connected layer with a ReLU activation function
network.add(layers.Dense(units=100, activation="relu"))
# Add fully connected layer with a softmax activation function
network.add(layers.Dense(units=46, activation="softmax"))
# Compile neural network
network.compile(loss="categorical_crossentropy", # Cross-entropy
optimizer="rmsprop", # Root Mean Square Propagation
metrics=["accuracy"]) # Accuracy performance metric
# Train neural network
history = network.fit(features_train, # Features
target_train, # Target
epochs=3, # Three epochs
verbose=0, # No output
batch_size=100, # Number of observations per batch
validation_data=(features_test, target_test)) # Test data
Using TensorFlow backend.
Discussion
In this solution we created a similar neural network to the binary classifier from the
last recipe, but with some notable changes. First, our data is 11,228 Reuters news‐
wires. Each newswire is categorized into 46 topics. We prepared our feature data by
converting the newswires into 5,000 binary features (denoting the presence of a cer‐
tain word in the newswires). We prepared the target data by one-hot encoding it so
that we obtain a target matrix denoting which of the 46 classes an observation
belongs to:
# View target matrix
target_train
array([[ 0.,
[ 0.,
[ 0.,
...,
[ 0.,
[ 0.,
[ 0.,
0.,
0.,
0.,
0., ...,
0., ...,
0., ...,
0.,
0.,
0.,
0.,
0.,
0.,
0.],
0.],
0.],
0.,
0.,
0.,
0., ...,
0., ...,
0., ...,
0.,
0.,
0.,
0.,
0.,
0.,
0.],
0.],
0.]])
Second, we increased the number of units in each of the hidden layers to help the
neural network represent the more complex relationship between the 46 classes.
306
|
Chapter 20: Neural Networks
Third, since this is a multiclass classification problem, we used an output layer with
46 units (one per class) containing a softmax activation function. The softmax activa‐
tion function will return an array of 46 values summing to 1. These 46 values repre‐
sent an observation’s probability of being a member of each of the 46 classes.
Fourth, we used a loss function suited to multiclass classification, the categorical
cross-entropy loss function, categorical_crossentropy.
20.5 Training a Regressor
Problem
You want to train a neural network for regression.
Solution
Use Keras to construct a feedforward neural network with a single output unit and no
activation function:
# Load libraries
import numpy as np
from keras.preprocessing.text import Tokenizer
from keras import models
from keras import layers
from sklearn.datasets import make_regression
from sklearn.model_selection import train_test_split
from sklearn import preprocessing
# Set random seed
np.random.seed(0)
# Generate features matrix and target vector
features, target = make_regression(n_samples = 10000,
n_features = 3,
n_informative = 3,
n_targets = 1,
noise = 0.0,
random_state = 0)
# Divide our data into training and test sets
features_train, features_test, target_train, target_test = train_test_split(
features, target, test_size=0.33, random_state=0)
# Start neural network
network = models.Sequential()
# Add fully connected layer with a ReLU activation function
network.add(layers.Dense(units=32,
activation="relu",
input_shape=(features_train.shape[1],)))
20.5 Training a Regressor
|
307
# Add fully connected layer with a ReLU activation function
network.add(layers.Dense(units=32, activation="relu"))
# Add fully connected layer with no activation function
network.add(layers.Dense(units=1))
# Compile neural network
network.compile(loss="mse", # Mean squared error
optimizer="RMSprop", # Optimization algorithm
metrics=["mse"]) # Mean squared error
# Train neural network
history = network.fit(features_train, # Features
target_train, # Target vector
epochs=10, # Number of epochs
verbose=0, # No output
batch_size=100, # Number of observations per batch
validation_data=(features_test, target_test)) # Test data
Using TensorFlow backend.
Discussion
It is completely possible to create a neural network to predict continuous values
instead of class probabilities. In the case of our binary classifier (Recipe 20.3) we used
an output layer with a single unit and a sigmoid activation function to produce a
probability that an observation was class 1. Importantly, the sigmoid activation func‐
tion constrained the outputted value to between 0 and 1. If we remove that constraint
by having no activation function, we allow the output to be a continuous value.
Furthermore, because we are training a regression, we should use an appropriate loss
function and evaluation metric, in our case the mean square error:
MSE =
1 n
yi − y i
ni∑
=1
2
where n is the number of observations; yi is the true value of the target we are trying
to predict, y, for observation i; and ŷi is the model’s predicted value for yi.
Finally, because we are using simulated data using scikit-learn, make_regression, we
didn’t have to standardize the features. It should be noted, however, that in almost all
real-world cases standardization would be necessary.
308
|
Chapter 20: Neural Networks
20.6 Making Predictions
Problem
You want to use a neural network to make predictions.
Solution
Use Keras to construct a feedforward neural network, then make predictions using
predict:
# Load libraries
import numpy as np
from keras.datasets import imdb
from keras.preprocessing.text import Tokenizer
from keras import models
from keras import layers
# Set random seed
np.random.seed(0)
# Set the number of features we want
number_of_features = 10000
# Load data and target vector from IMDB movie data
(data_train, target_train), (data_test, target_test) = imdb.load_data(
num_words=number_of_features)
# Convert IMDB data to a one-hot encoded feature matrix
tokenizer = Tokenizer(num_words=number_of_features)
features_train = tokenizer.sequences_to_matrix(data_train, mode="binary")
features_test = tokenizer.sequences_to_matrix(data_test, mode="binary")
# Start neural network
network = models.Sequential()
# Add fully connected layer with a ReLU activation function
network.add(layers.Dense(units=16,
activation="relu",
input_shape=(number_of_features,)))
# Add fully connected layer with a ReLU activation function
network.add(layers.Dense(units=16, activation="relu"))
# Add fully connected layer with a sigmoid activation function
network.add(layers.Dense(units=1, activation="sigmoid"))
# Compile neural network
network.compile(loss="binary_crossentropy", # Cross-entropy
optimizer="rmsprop", # Root Mean Square Propagation
metrics=["accuracy"]) # Accuracy performance metric
20.6 Making Predictions
|
309
# Train neural network
history = network.fit(features_train, # Features
target_train, # Target vector
epochs=3, # Number of epochs
verbose=0, # No output
batch_size=100, # Number of observations per batch
validation_data=(features_test, target_test)) # Test data
# Predict classes of test set
predicted_target = network.predict(features_test)
Using TensorFlow backend.
Discussion
Making predictions is easy in Keras. Once we have trained our neural network we can
use the predict method, which takes as an argument a set of features and returns the
predicted output for each observation. In our solution our neural network is set up
for binary classification so the predicted output is the probability of being class 1.
Observations with predicted values very close to 1 are highly likely to be class 1, while
observations with predicted values very close to 0 are highly likely to be class 0. For
example, this is the predicted probability that the first observation in our test feature
matrix is class 1:
# View the probability the first observation is class 1
predicted_target[0]
array([ 0.83937484], dtype=float32)
20.7 Visualize Training History
Problem
You want to find the “sweet spot” in a neural network’s loss and/or accuracy score.
Solution
Use Matplotlib to visualize the loss of the test and training set over each epoch:
# Load libraries
import numpy as np
from keras.datasets import imdb
from keras.preprocessing.text import Tokenizer
from keras import models
from keras import layers
import matplotlib.pyplot as plt
# Set random seed
np.random.seed(0)
310
|
Chapter 20: Neural Networks
# Set the number of features we want
number_of_features = 10000
# Load data and target vector from movie review data
(data_train, target_train), (data_test, target_test) = imdb.load_data(
num_words=number_of_features)
# Convert movie review data to a one-hot encoded feature matrix
tokenizer = Tokenizer(num_words=number_of_features)
features_train = tokenizer.sequences_to_matrix(data_train, mode="binary")
features_test = tokenizer.sequences_to_matrix(data_test, mode="binary")
# Start neural network
network = models.Sequential()
# Add fully connected layer with a ReLU activation function
network.add(layers.Dense(units=16,
activation="relu",
input_shape=(number_of_features,)))
# Add fully connected layer with a ReLU activation function
network.add(layers.Dense(units=16, activation="relu"))
# Add fully connected layer with a sigmoid activation function
network.add(layers.Dense(units=1, activation="sigmoid"))
# Compile neural network
network.compile(loss="binary_crossentropy", # Cross-entropy
optimizer="rmsprop", # Root Mean Square Propagation
metrics=["accuracy"]) # Accuracy performance metric
# Train neural network
history = network.fit(features_train, # Features
target_train, # Target
epochs=15, # Number of epochs
verbose=0, # No output
batch_size=1000, # Number of observations per batch
validation_data=(features_test, target_test)) # Test data
# Get training and test loss histories
training_loss = history.history["loss"]
test_loss = history.history["val_loss"]
# Create count of the number of epochs
epoch_count = range(1, len(training_loss) + 1)
# Visualize loss history
plt.plot(epoch_count, training_loss, "r--")
plt.plot(epoch_count, test_loss, "b-")
plt.legend(["Training Loss", "Test Loss"])
plt.xlabel("Epoch")
20.7 Visualize Training History
|
311
plt.ylabel("Loss")
plt.show();
Using TensorFlow backend.
Alternatively, we can use the same approach to visualize the training and test accu‐
racy over each epoch:
# Get training and test accuracy histories
training_accuracy = history.history["acc"]
test_accuracy = history.history["val_acc"]
plt.plot(epoch_count, training_accuracy, "r--")
plt.plot(epoch_count, test_accuracy, "b-")
# Visualize accuracy history
plt.legend(["Training Accuracy", "Test Accuracy"])
plt.xlabel("Epoch")
plt.ylabel("Accuracy Score")
plt.show();
312
|
Chapter 20: Neural Networks
Discussion
When our neural network is new, it will have a poor performance. As the neural net‐
work learns on the training data, the model’s error on both the training and test set
will tend to increase. However, at a certain point the neural network starts “memoriz‐
ing” the training data, and overfits. When this starts happening, the training error
will decrease while the test error will start increasing. Therefore, in many cases there
is a “sweet spot” where the test error (which is the error we mainly care about) is at its
lowest point. This effect can be plainly seen in the solution where we visualize the
training and test loss at each epoch. Note that the test error is lowest around epoch
five, after which the training loss continues to increase while the test loss starts
increasing. At this point onward, the model is overfitting.
20.8 Reducing Overfitting with Weight Regularization
Problem
You want to reduce overfitting.
Solution
Try penalizing the parameters of the network, also called weight regularization:
20.8 Reducing Overfitting with Weight Regularization
|
313
# Load libraries
import numpy as np
from keras.datasets import imdb
from keras.preprocessing.text import Tokenizer
from keras import models
from keras import layers
from keras import regularizers
# Set random seed
np.random.seed(0)
# Set the number of features we want
number_of_features = 1000
# Load data and target vector from movie review data
(data_train, target_train), (data_test, target_test) = imdb.load_data(
num_words=number_of_features)
# Convert movie review data to a one-hot encoded feature matrix
tokenizer = Tokenizer(num_words=number_of_features)
features_train = tokenizer.sequences_to_matrix(data_train, mode="binary")
features_test = tokenizer.sequences_to_matrix(data_test, mode="binary")
# Start neural network
network = models.Sequential()
# Add fully connected layer with a ReLU activation function
network.add(layers.Dense(units=16,
activation="relu",
kernel_regularizer=regularizers.l2(0.01),
input_shape=(number_of_features,)))
# Add fully connected layer with a ReLU activation function
network.add(layers.Dense(units=16,
kernel_regularizer=regularizers.l2(0.01),
activation="relu"))
# Add fully connected layer with a sigmoid activation function
network.add(layers.Dense(units=1, activation="sigmoid"))
# Compile neural network
network.compile(loss="binary_crossentropy", # Cross-entropy
optimizer="rmsprop", # Root Mean Square Propagation
metrics=["accuracy"]) # Accuracy performance metric
# Train neural network
history = network.fit(features_train, # Features
target_train, # Target vector
epochs=3, # Number of epochs
verbose=0, # No output
batch_size=100, # Number of observations per batch
validation_data=(features_test, target_test)) # Test data
314
|
Chapter 20: Neural Networks
Using TensorFlow backend.
Discussion
One strategy to combat overfitting neural networks is by penalizing the parameters
(i.e., weights) of the neural network such that they are driven to be small values—cre‐
ating a simpler model less prone to overfit. This method is called weight regulariza‐
tion or weight decay. More specifically, in weight regularization a penalty is added to
the loss function, such as the L2 norm.
In Keras, we can add a weight regularization by including using kernel_regular
izer=regularizers.l2(0.01) in a layer’s parameters. In this example, 0.01 deter‐
mines how much we penalize higher parameter values.
20.9 Reducing Overfitting with Early Stopping
Problem
You want to reduce overfitting.
Solution
Try stopping training when the test loss stops decreasing, a strategy called early stop‐
ping:
# Load libraries
import numpy as np
from keras.datasets import imdb
from keras.preprocessing.text import Tokenizer
from keras import models
from keras import layers
from keras.callbacks import EarlyStopping, ModelCheckpoint
# Set random seed
np.random.seed(0)
# Set the number of features we want
number_of_features = 1000
# Load data and target vector from movie review data
(data_train, target_train), (data_test, target_test) = imdb.load_data(
num_words=number_of_features)
# Convert movie review data to a one-hot encoded feature matrix
tokenizer = Tokenizer(num_words=number_of_features)
features_train = tokenizer.sequences_to_matrix(data_train, mode="binary")
features_test = tokenizer.sequences_to_matrix(data_test, mode="binary")
# Start neural network
20.9 Reducing Overfitting with Early Stopping
|
315
network = models.Sequential()
# Add fully connected layer with a ReLU activation function
network.add(layers.Dense(units=16,
activation="relu",
input_shape=(number_of_features,)))
# Add fully connected layer with a ReLU activation function
network.add(layers.Dense(units=16, activation="relu"))
# Add fully connected layer with a sigmoid activation function
network.add(layers.Dense(units=1, activation="sigmoid"))
# Compile neural network
network.compile(loss="binary_crossentropy", # Cross-entropy
optimizer="rmsprop", # Root Mean Square Propagation
metrics=["accuracy"]) # Accuracy performance metric
# Set callback functions to early stop training and save the best model so far
callbacks = [EarlyStopping(monitor="val_loss", patience=2),
ModelCheckpoint(filepath="best_model.h5",
monitor="val_loss",
save_best_only=True)]
# Train neural network
history = network.fit(features_train, # Features
target_train, # Target vector
epochs=20, # Number of epochs
callbacks=callbacks, # Early stopping
verbose=0, # Print description after each epoch
batch_size=100, # Number of observations per batch
validation_data=(features_test, target_test)) # Test data
Using TensorFlow backend.
Discussion
As we discussed in Recipe 20.7, typically in the first training epochs both the training
and test errors will decrease, but at some point the network will start “memorizing”
the training data, causing the training error to continue to decrease even while the
test error starts increasing. Because of this phenomenon, one of the most common
and very effective methods to counter overfitting is to monitor the training process
and stop training when the test error starts to increase. This strategy is called early
stopping.
In Keras, we can implement early stopping as a callback function. Callbacks are func‐
tions that can be applied at certain stages of the training process, such as at the end of
each epoch. Specifically, in our solution, we included EarlyStopping(moni
tor='val_loss', patience=2) to define that we wanted to monitor the test (valida‐
tion) loss at each epoch and after the test loss has not improved after two epochs,
316
|
Chapter 20: Neural Networks
training is interrupted. However, since we set patience=2, we won’t get the best
model, but the model two epochs after the best model. Therefore, optionally, we can
include a second operation, ModelCheckpoint, which saves the model to a file after
every checkpoint (which can be useful in case a multiday training session is interrup‐
ted for some reason). It would be helpful for us, if we set save_best_only=True,
because then ModelCheckpoint will only save the best model.
20.10 Reducing Overfitting with Dropout
Problem
You want to reduce overfitting.
Solution
Introduce noise into your network’s architecture using dropout:
# Load libraries
import numpy as np
from keras.datasets import imdb
from keras.preprocessing.text import Tokenizer
from keras import models
from keras import layers
# Set random seed
np.random.seed(0)
# Set the number of features we want
number_of_features = 1000
# Load data and target vector from movie review data
(data_train, target_train), (data_test, target_test) = imdb.load_data(
num_words=number_of_features)
# Convert movie review data to a one-hot encoded feature matrix
tokenizer = Tokenizer(num_words=number_of_features)
features_train = tokenizer.sequences_to_matrix(data_train, mode="binary")
features_test = tokenizer.sequences_to_matrix(data_test, mode="binary")
# Start neural network
network = models.Sequential()
# Add a dropout layer for input layer
network.add(layers.Dropout(0.2, input_shape=(number_of_features,)))
# Add fully connected layer with a ReLU activation function
network.add(layers.Dense(units=16, activation="relu"))
# Add a dropout layer for previous hidden layer
20.10 Reducing Overfitting with Dropout
|
317
network.add(layers.Dropout(0.5))
# Add fully connected layer with a ReLU activation function
network.add(layers.Dense(units=16, activation="relu"))
# Add a dropout layer for previous hidden layer
network.add(layers.Dropout(0.5))
# Add fully connected layer with a sigmoid activation function
network.add(layers.Dense(units=1, activation="sigmoid"))
# Compile neural network
network.compile(loss="binary_crossentropy", # Cross-entropy
optimizer="rmsprop", # Root Mean Square Propagation
metrics=["accuracy"]) # Accuracy performance metric
# Train neural network
history = network.fit(features_train, # Features
target_train, # Target vector
epochs=3, # Number of epochs
verbose=0, # No output
batch_size=100, # Number of observations per batch
validation_data=(features_test, target_test)) # Test data
Using TensorFlow backend.
Discussion
Dropout is a popular and powerful method for regularizing neural networks. In
dropout, every time a batch of observations is created for training, a proportion of
the units in one or more layers is multiplied by zero (i.e., dropped). In this setting,
every batch is trained on the same network (e.g., the same parameters), but each
batch is confronted by a slightly different version of that network’s architecture.
Dropout is effective because by constantly and randomly dropping units in each
batch, it forces units to learn parameter values able to perform under a wide variety
of network architectures. That is, they learn to be robust to disruptions (i.e., noise) in
the other hidden units, and this prevents the network from simply memorizing the
training data.
It is possible to add dropout to both the hidden and input layers. When an input layer
is dropped, its feature value is not introduced into the network for that batch. A com‐
mon choice for the portion of units to drop is 0.2 for input units and 0.5 for hidden
units.
In Keras, we can implement dropout by adding Dropout layers into our network
architecture. Each Dropout layer will drop a user-defined hyperparameter of units in
the previous layer every batch. Remember in Keras the input layer is assumed to be
the first layer and not added using add. Therefore, if we want to add dropout to the
318
|
Chapter 20: Neural Networks
input layer, the first layer we add in our network architecture is a dropout layer. This
layer contains both the proportion of the input layer’s units to drop 0.2 and
input_shape defining the shape of the observation data. Next, we add a dropout layer
with 0.5 after each of the hidden layers.
20.11 Saving Model Training Progress
Problem
Given a neural network that will take a long time to train, you want to save your pro‐
gress in case the training process is interrupted.
Solution
Use the callback function ModelCheckpoint to save the model after every epoch:
# Load libraries
import numpy as np
from keras.datasets import imdb
from keras.preprocessing.text import Tokenizer
from keras import models
from keras import layers
from keras.callbacks import ModelCheckpoint
# Set random seed
np.random.seed(0)
# Set the number of features we want
number_of_features = 1000
# Load data and target vector from movie review data
(data_train, target_train), (data_test, target_test) = imdb.load_data(
num_words=number_of_features)
# Convert movie review data to a one-hot encoded feature matrix
tokenizer = Tokenizer(num_words=number_of_features)
features_train = tokenizer.sequences_to_matrix(data_train, mode="binary")
features_test = tokenizer.sequences_to_matrix(data_test, mode="binary")
# Start neural network
network = models.Sequential()
# Add fully connected layer with a ReLU activation function
network.add(layers.Dense(units=16,
activation="relu",
input_shape=(number_of_features,)))
# Add fully connected layer with a ReLU activation function
network.add(layers.Dense(units=16, activation="relu"))
20.11 Saving Model Training Progress
|
319
# Add fully connected layer with a sigmoid activation function
network.add(layers.Dense(units=1, activation="sigmoid"))
# Compile neural network
network.compile(loss="binary_crossentropy", # Cross-entropy
optimizer="rmsprop", # Root Mean Square Propagation
metrics=["accuracy"]) # Accuracy performance metric
# Set callback functions to early stop training and save the best model so far
checkpoint = [ModelCheckpoint(filepath="models.hdf5")]
# Train neural network
history = network.fit(features_train, # Features
target_train, # Target vector
epochs=3, # Number of epochs
callbacks=checkpoint, # Checkpoint
verbose=0, # No output
batch_size=100, # Number of observations per batch
validation_data=(features_test, target_test)) # Test data
Using TensorFlow backend.
Discussion
In Recipe 20.8 we used the callback function ModelCheckpoint in conjunction with
EarlyStopping to end monitoring and end training when the test error stopped
improving. However, there is another, more mundane reason for using ModelCheck
point. In the real world, it is common for neural networks to train for hours or even
days. During that time a lot can go wrong: computers can lose power, servers can
crash, or inconsiderate graduate students can close your laptop.
ModelCheckpoint alleviates this problem by saving the model after every epoch.
Specifically, after every epoch ModelCheckpoint saves a model to the location speci‐
fied by the filepath parameter. If we include only a filename (e.g., models.hdf5) that
file will be overridden with the latest model every epoch. If we only wanted to save
the best model according to the performance of some loss function, we can set
save_best_only=True and monitor='val_loss' to not override a file if the model
has a worse test loss than the previous model. Alternatively, we can save every epoch’s
model as its own file by including the epoch number and test loss score into the file‐
name itself. For example, if we set filepath to model_{epoch:02d}_{val_loss:.
2f}.hdf5, the name of the file containing the model saved after the 11th epoch with a
test loss value of 0.33 would be model_10_0.35.hdf5 (notice that the epoch number is
0-indexed).
320
|
Chapter 20: Neural Networks
20.12 k-Fold Cross-Validating Neural Networks
Problem
You want to evaluate a neural network using k-fold cross-validation.
Solution
Often k-fold cross-validating neural networks is neither necessary nor advisable.
However, if it is appropriate: use Keras’ scikit-learn wrapper to allow Keras’ sequential
models to use the scikit-learn API:
# Load libraries
import numpy as np
from keras import models
from keras import layers
from keras.wrappers.scikit_learn import KerasClassifier
from sklearn.model_selection import cross_val_score
from sklearn.datasets import make_classification
# Set random seed
np.random.seed(0)
# Number of features
number_of_features = 100
# Generate features matrix and target vector
features, target = make_classification(n_samples = 10000,
n_features = number_of_features,
n_informative = 3,
n_redundant = 0,
n_classes = 2,
weights = [.5, .5],
random_state = 0)
# Create function returning a compiled network
def create_network():
# Start neural network
network = models.Sequential()
# Add fully connected layer with a ReLU activation function
network.add(layers.Dense(units=16, activation="relu", input_shape=(
number_of_features,)))
# Add fully connected layer with a ReLU activation function
network.add(layers.Dense(units=16, activation="relu"))
# Add fully connected layer with a sigmoid activation function
network.add(layers.Dense(units=1, activation="sigmoid"))
20.12 k-Fold Cross-Validating Neural Networks
|
321
# Compile neural network
network.compile(loss="binary_crossentropy", # Cross-entropy
optimizer="rmsprop", # Root Mean Square Propagation
metrics=["accuracy"]) # Accuracy performance metric
# Return compiled network
return network
# Wrap Keras model so it can be used by scikit-learn
neural_network = KerasClassifier(build_fn=create_network,
epochs=10,
batch_size=100,
verbose=0)
# Evaluate neural network using three-fold cross-validation
cross_val_score(neural_network, features, target, cv=3)
Using TensorFlow backend.
array([ 0.90461907,
0.77437743,
0.87068707])
Discussion
Theoretically, there is no reason we cannot use cross-validation to evaluate neural
networks. However, neural networks are often used on very large data and can take
hours or even days to train. For this reason, if the training time is long, adding the
computational expense of k-fold cross-validation is unadvisable. For example, a
model normally taking one day to train would take 10 days to evaluate using 10-fold
cross-validation. If we have large data, it is often appropriate to simply evaluate the
neural network on some test set.
If we have smaller data, k-fold cross-validation can be useful to maximize our ability
to evaluate the neural network’s performance. This is possible in Keras because we
can “wrap” any neural network such that it can use the evaluation features available in
scikit-learn, including k-fold cross-validation. To accomplish this, we first have to
create a function that returns a compiled neural network. Next we use KerasClassi
fier (if we have a classifier; if we have a regressor we can use KerasRegressor) to
wrap the model so it can be used by scikit-learn. After this, we can use our neural
network like any other scikit-learn learning algorithm (e.g., random forests, logistic
regression). In our solution, we used cross_val_score to run a three-fold crossvalidation on our neural network.
20.13 Tuning Neural Networks
Problem
You want to automatically select the best hyperparameters for your neural network.
322
|
Chapter 20: Neural Networks
Solution
Combine a Keras neural network with scikit-learn’s model selection tools like Grid
SearchCV:
# Load libraries
import numpy as np
from keras import models
from keras import layers
from keras.wrappers.scikit_learn import KerasClassifier
from sklearn.model_selection import GridSearchCV
from sklearn.datasets import make_classification
# Set random seed
np.random.seed(0)
# Number of features
number_of_features = 100
# Generate features matrix and target vector
features, target = make_classification(n_samples = 10000,
n_features = number_of_features,
n_informative = 3,
n_redundant = 0,
n_classes = 2,
weights = [.5, .5],
random_state = 0)
# Create function returning a compiled network
def create_network(optimizer="rmsprop"):
# Start neural network
network = models.Sequential()
# Add fully connected layer with a ReLU activation function
network.add(layers.Dense(units=16,
activation="relu",
input_shape=(number_of_features,)))
# Add fully connected layer with a ReLU activation function
network.add(layers.Dense(units=16, activation="relu"))
# Add fully connected layer with a sigmoid activation function
network.add(layers.Dense(units=1, activation="sigmoid"))
# Compile neural network
network.compile(loss="binary_crossentropy", # Cross-entropy
optimizer=optimizer, # Optimizer
metrics=["accuracy"]) # Accuracy performance metric
# Return compiled network
return network
20.13 Tuning Neural Networks
|
323
# Wrap Keras model so it can be used by scikit-learn
neural_network = KerasClassifier(build_fn=create_network, verbose=0)
# Create hyperparameter space
epochs = [5, 10]
batches = [5, 10, 100]
optimizers = ["rmsprop", "adam"]
# Create hyperparameter options
hyperparameters = dict(optimizer=optimizers, epochs=epochs, batch_size=batches)
# Create grid search
grid = GridSearchCV(estimator=neural_network, param_grid=hyperparameters)
# Fit grid search
grid_result = grid.fit(features, target)
Using TensorFlow backend.
Discussion
In Recipes 12.1 and 12.2, we covered using scikit-learn’s model selection techniques
to identify the best hyperparameters of a scikit-learn model. In Recipe 20.12 we
learned that we can wrap our neural network so it can use the scikit-learn API. In this
recipe we combine these two techniques to identify the best hyperparameters of a
neural network.
The hyperparameters of a model are important and should be selected with care.
However, before we get it into our heads that model selection strategies like grid
search are a good idea, we must realize that if our model would normally have taken
12 hours or a day to train, this grid search process could take a week or more. There‐
fore, automatic hyperparameter tuning of neural networks is not the silver bullet, but
it is a useful tool to have in certain circumstances.
In our solution we conducted a cross-validated grid search over a number of options
for the optimization algorithm, number of epochs, and batch size. Even this toy
example took a few minutes to run, but once it is done we can use best_params_ to
view the hyperparameters of the neural network with the best results:
# View hyperparameters of best neural network
grid_result.best_params_
{'batch_size': 10, 'epochs': 5, 'optimizer': 'adam'}
324
|
Chapter 20: Neural Networks
20.14 Visualizing Neural Networks
Problem
You want to quickly visualize a neural network’s architecture.
Solution
Use Keras’ model_to_dot or plot_model:
# Load libraries
from keras import models
from keras import layers
from IPython.display import SVG
from keras.utils.vis_utils import model_to_dot
from keras.utils import plot_model
# Start neural network
network = models.Sequential()
# Add fully connected layer with a ReLU activation function
network.add(layers.Dense(units=16, activation="relu", input_shape=(10,)))
# Add fully connected layer with a ReLU activation function
network.add(layers.Dense(units=16, activation="relu"))
# Add fully connected layer with a sigmoid activation function
network.add(layers.Dense(units=1, activation="sigmoid"))
# Visualize network architecture
SVG(model_to_dot(network, show_shapes=True).create(prog="dot", format="svg"))
Using TensorFlow backend.
20.14 Visualizing Neural Networks
|
325
Alternatively, if we want to save the visualization as a file, we can use plot_model:
# Save the visualization as a file
plot_model(network, show_shapes=True, to_file="network.png")
Discussion
Keras provides utility functions to quickly visualize neural networks. If we wish to
display a neural network in a Jupyter Notebook, we can use model_to_dot. The
show_shapes parameter shows the shape of the inputs and outputs and can help with
debugging. For a simpler model, we can set show_shapes=True:
# Visualize network architecture
SVG(model_to_dot(network, show_shapes=False).create(prog="dot", format="svg"))
326
|
Chapter 20: Neural Networks
20.15 Classifying Images
Problem
You want to classify images using a convolutional neural network.
Solution
Use Keras to create a neural network with at least one convolutional layer:
import numpy as np
from keras.datasets import mnist
from keras.models import Sequential
from keras.layers import Dense, Dropout, Flatten
from keras.layers.convolutional import Conv2D, MaxPooling2D
from keras.utils import np_utils
from keras import backend as K
# Set that the color channel value will be first
K.set_image_data_format("channels_first")
# Set seed
np.random.seed(0)
20.15 Classifying Images
|
327
# Set image information
channels = 1
height = 28
width = 28
# Load data and target from MNIST data
(data_train, target_train), (data_test, target_test) = mnist.load_data()
# Reshape training image data into features
data_train = data_train.reshape(data_train.shape[0], channels, height, width)
# Reshape test image data into features
data_test = data_test.reshape(data_test.shape[0], channels, height, width)
# Rescale pixel intensity to between 0 and 1
features_train = data_train / 255
features_test = data_test / 255
# One-hot encode target
target_train = np_utils.to_categorical(target_train)
target_test = np_utils.to_categorical(target_test)
number_of_classes = target_test.shape[1]
# Start neural network
network = Sequential()
# Add convolutional layer with 64 filters, a 5x5 window, and ReLU activation function
network.add(Conv2D(filters=64,
kernel_size=(5, 5),
input_shape=(channels, width, height),
activation='relu'))
# Add max pooling layer with a 2x2 window
network.add(MaxPooling2D(pool_size=(2, 2)))
# Add dropout layer
network.add(Dropout(0.5))
# Add layer to flatten input
network.add(Flatten())
# # Add fully connected layer of 128 units with a ReLU activation function
network.add(Dense(128, activation="relu"))
# Add dropout layer
network.add(Dropout(0.5))
# Add fully connected layer with a softmax activation function
network.add(Dense(number_of_classes, activation="softmax"))
# Compile neural network
328
|
Chapter 20: Neural Networks
network.compile(loss="categorical_crossentropy", # Cross-entropy
optimizer="rmsprop", # Root Mean Square Propagation
metrics=["accuracy"]) # Accuracy performance metric
# Train neural network
network.fit(features_train, # Features
target_train, # Target
epochs=2, # Number of epochs
verbose=0, # Don't print description after each epoch
batch_size=1000, # Number of observations per batch
validation_data=(features_test, target_test)) # Data for evaluation
Using TensorFlow backend.
<keras.callbacks.History at 0x133f37e80>
Discussion
Convolutional neural networks (also called ConvNets) are a popular type of network
that has proven very effective at computer vision (e.g., recognizing cats, dogs, planes,
and even hot dogs). It is completely possible to use feedforward neural networks on
images, where each pixel is a feature. However, when doing so we run into two major
problems. First, feedforward neural networks do not take into account the spatial
structure of the pixels. For example, in a 10 × 10 pixel image we might convert it into
a vector of 100 pixel features, and in this case feedforward would consider the first
feature (e.g., pixel value) to have the same relationship with the 10th feature as the
11th feature. However, in reality the 10th feature represents a pixel on the far side of
the image as the first feature, while the 11th feature represents the pixel immediately
below the first pixel. Second, and relatedly, feedforward neural networks learn global
relationships in the features instead of local patterns. In more practical terms, this
means that feedforward neural networks are not able to detect an object regardless of
where it appears in an image. For example, imagine we are training a neural network
to recognize faces, and these faces might appear anywhere in the image from the
upper right to the middle to the lower left.
The power of convolutional neural networks is their ability handle both of these
issues (and others). A complete explanation of convolutional neural networks is well
beyond the scope of this book, but a brief explanation will be useful. The data of an
individual image contains two or three dimensions: height, width, and depth. The
first two should be obvious, but the last deserves explanation. The depth is the color
of a pixel. In grayscale images the depth is only one (the intensity of the pixel) and
therefore the image is a matrix. However, in color images a pixel’s color is represented
by multiple values. For example, in an RGB image a pixel’s color is represented by
three values representing red, green, and blue. Therefore, the data for an image can
be imagined to be a three-dimensional tensor: width × height × depth (called feature
maps). In convolutional neural networks, a convolution (don’t worry if you don’t
20.15 Classifying Images
|
329
know what that means) can be imagined to slide a window over the pixels of an
image, looking at both the individual pixel and also its neighbors. It then transforms
the raw image data into a new three-dimensional tensor where the first two dimen‐
sions are approximately width and height, while the third dimension (which con‐
tained the color values) now represents the patterns—called filters—(for example, a
sharp corner or sweeping gradient) to which that pixel “belongs.”
The second important concept for our purposes is that of pooling layers. Pooling lay‐
ers move a window over our data (though usually they only look at every nth pixel,
called striding) downsizing our data by summarizing the window in some way. The
most common method is max pooling where the maximum value in each window is
sent to the next layer. One reason for max pooling is merely practical; the convolu‐
tional process creates many more parameters to learn, which can get ungainly very
quickly. Second, more intuitively, max pooling can be thought of as “zooming out” of
an image.
An example might be useful here. Imagine we have an image containing a dog’s face.
The first convolutional layer might find patterns like shape edges. Then we use a max
pool layer to “zoom out,” and a second convolutional layer to find patterns like dog
ears. Finally we use another max pooling layer to zoom out again and a final convolu‐
tional layer to find patterns like dogs’ faces.
Finally, fully connected layers are often used at the end of the network to do the
actual classification.
While our solution might look like a lot of lines of code, it is actually very similar to
our binary classifier from earlier in this chapter. In this solution we used the famous
MNIST dataset, which is a de facto benchmark dataset in machine learning. The
MNIST dataset contains 70,000 small images (28 × 28) of handwritten digits from 0
to 9. This dataset is labeled so that we know the actual digit (i.e., class) for each small
image. The standard training-test split is to use 60,000 images for training and 10,000
for testing.
We reorganized the data into the format expected by a convolutional network. Specif‐
ically, we used reshape to convert the observation data such that it is the shape Keras
expects. Next, we rescaled the values to be between 0 and 1, since training perfor‐
mance can suffer if an observation’s values are much greater than the network’s
parameters (which are initialized as small numbers). Finally, we one-hot encoded the
target data so that every observation’s target has 10 classes, representing the digits
0–9.
With the image data process we can build our convolutional network. First, we add a
convolutional layer and specify the number of filters and other characteristics. The
size of the window is a hyperparameter; however, 3 × 3 is standard practice for most
images, while larger windows are often used in larger images. Second, we add a max
330
|
Chapter 20: Neural Networks
pooling layer, summarizing the nearby pixels. Third, we add a dropout layer to
reduce the chances of overfitting. Fourth, we add a flatten layer to convert the convo‐
lutionary inputs into a format able to be used by a fully connected layer. Finally, we
add the fully connected layers and an output layer to do the actual classification.
Notice that because this is a multiclass classification problem, we use the softmax
activation function in the output layer.
It should be noted that this is a pretty simple convolutional neural network. It is com‐
mon to see a much deeper network with many more convolutional and max pooling
layers stacked together.
20.16 Improving Performance with Image Augmentation
Problem
You want to improve the performance of your convolutional neural network.
Solution
For better results, preprocess the images and augment the data beforehand using
ImageDataGenerator:
# Load library
from keras.preprocessing.image import ImageDataGenerator
# Create image augmentation
augmentation = ImageDataGenerator(featurewise_center=True, # Apply ZCA whitening
zoom_range=0.3, # Randomly zoom in on images
width_shift_range=0.2, # Randomly shift images
horizontal_flip=True, # Randomly flip images
rotation_range=90) # Randomly rotate
# Process all images from the directory 'raw/images'
augment_images = augmentation.flow_from_directory("raw/images", # Image folder
batch_size=32, # Batch size
class_mode="binary", # Classes
save_to_dir="processed/images")
Using TensorFlow backend.
Found 12665 images belonging to 2 classes.
Discussion
First an apology—this solution’s code will not immediately run for you because you
do not have the required folders of images. However, since the most common situa‐
20.16 Improving Performance with Image Augmentation
|
331
tion is having a directory of images, I wanted to include it. The technique should
easily translate to your own images.
One way to improve the performance of a convolutional neural network is to prepro‐
cess the images. We discussed a number of techniques in Chapter 8; however, it is
worth noting that Keras’ ImageDataGenerator contains a number of basic prepro‐
cessing techniques. For example, in our solution we used featurewise_center=True
to standardize the pixels across the entire data.
A second technique to improve performance is to add noise. An interesting feature of
neural networks is that, counterintuitively, their performance often improves when
noise is added to the data. The reason is that the additional noise can make the neural
networks more robust in the face of real-world noise and prevents them from overfit‐
ting the data.
When training convolutional neural networks for images, we can add noise to our
observations by randomly transforming images in various ways, such as flipping
images or zooming in on images. Even small changes can significantly improve
model performance. We can use the same ImageDataGenerator class to conduct
these transformations. The Keras documentation (referenced in “See Also” on page
333) specifies the full list of available transformations; however, our example contains
a sample of them are, including random zooming, shifting, flipping, and rotation.
It is important to note that the output of flow_from_directory is a Python generator
object. This is because in most cases we will want to process the images on-demand
as they are sent to the neural network for training. If we wanted to process all the
images prior to training, we could simply iterate over the generator.
Finally, since augment_images is a generator, when training our neural network we
will have to use fit_generator instead of fit. For example:
# Train neural network
network.fit_generator(augment_images,
#Number of times to call the generator for each epoch
steps_per_epoch=2000,
# Number of epochs
epochs=5,
# Test data generator
validation_data=augment_images_test,
# Number of items to call the generator
# for each test epoch
validation_steps=800)
Note all the raw images used in this recipe are available on GitHub.
332
|
Chapter 20: Neural Networks
See Also
• Image Preprocessing, Keras
20.17 Classifying Text
Problem
You want to classify text data.
Solution
Use a long short-term memory recurrent neural network:
# Load libraries
import numpy as np
from keras.datasets import imdb
from keras.preprocessing import sequence
from keras import models
from keras import layers
# Set random seed
np.random.seed(0)
# Set the number of features we want
number_of_features = 1000
# Load data and target vector from movie review data
(data_train, target_train), (data_test, target_test) = imdb.load_data(
num_words=number_of_features)
# Use padding or truncation to make each observation have 400 features
features_train = sequence.pad_sequences(data_train, maxlen=400)
features_test = sequence.pad_sequences(data_test, maxlen=400)
# Start neural network
network = models.Sequential()
# Add an embedding layer
network.add(layers.Embedding(input_dim=number_of_features, output_dim=128))
# Add a long short-term memory layer with 128 units
network.add(layers.LSTM(units=128))
# Add fully connected layer with a sigmoid activation function
network.add(layers.Dense(units=1, activation="sigmoid"))
# Compile neural network
network.compile(loss="binary_crossentropy", # Cross-entropy
20.17 Classifying Text
|
333
optimizer="Adam", # Adam optimization
metrics=["accuracy"]) # Accuracy performance metric
# Train neural network
history = network.fit(features_train, # Features
target_train, # Target
epochs=3, # Number of epochs
verbose=0, # Do not print description after each epoch
batch_size=1000, # Number of observations per batch
validation_data=(features_test, target_test)) # Test data
Using TensorFlow backend.
Discussion
Oftentimes we have text data that we want to classify. While it is possible to use a type
of convolutional network, we are going to focus on a more popular option: the recur‐
rent neural network. The key feature of recurrent neural networks is that information
loops back in the network. This gives recurrent neural networks a type of memory
they can use to better understand sequential data. A popular type of recurrent neural
network is the long short-term memory (LSTM) network, which allows for informa‐
tion to loop backward in the network. For a more detailed explanation, see the addi‐
tional resources.
In this solution, we have our movie review data from Recipe 20.3 and we want to
train an LSTM network to predict if these reviews are positive or negative. Before we
can train our network, a little data processing is needed. Our text data comes in the
form of a list of integers:
# View first observation
print(data_train[0])
[1, 14, 22, 16, 43, 530, 973, 2, 2, 65, 458, 2, 66, 2, 4, 173, 36, 256, 5, 25,
100, 43, 838, 112, 50, 670, 2, 9, 35, 480, 284, 5, 150, 4, 172, 112, 167, 2,
336, 385, 39, 4, 172, 2, 2, 17, 546, 38, 13, 447, 4, 192, 50, 16, 6, 147, 2,
19, 14, 22, 4, 2, 2, 469, 4, 22, 71, 87, 12, 16, 43, 530, 38, 76, 15, 13, 2,
4, 22, 17, 515, 17, 12, 16, 626, 18, 2, 5, 62, 386, 12, 8, 316, 8, 106, 5, 4,
2, 2, 16, 480, 66, 2, 33, 4, 130, 12, 16, 38, 619, 5, 25, 124, 51, 36, 135,
48, 25, 2, 33, 6, 22, 12, 215, 28, 77, 52, 5, 14, 407, 16, 82, 2, 8, 4, 107,
117, 2, 15, 256, 4, 2, 7, 2, 5, 723, 36, 71, 43, 530, 476, 26, 400, 317, 46,
7, 4, 2, 2, 13, 104, 88, 4, 381, 15, 297, 98, 32, 2, 56, 26, 141, 6, 194, 2,
18, 4, 226, 22, 21, 134, 476, 26, 480, 5, 144, 30, 2, 18, 51, 36, 28, 224,
92, 25, 104, 4, 226, 65, 16, 38, 2, 88, 12, 16, 283, 5, 16, 2, 113, 103, 32,
15, 16, 2, 19, 178, 32]
Each integer in this list corresponds to some word. However, because each review
does not contain the same number of words, each observation is not the same length.
Therefore, before we can input this data into our neural network, we need to make all
the observations the same length. We can do this using pad_sequences. pad_sequen
334
|
Chapter 20: Neural Networks
ces pads each observation’s data so that they are all the same size. We can see this if
we look at our first observation after it is processed by pad_sequences:
# View first observation
print(features_test[0])
[
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
760
4
2
8 14 991
20 131
2
77
2 18
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
1
7
4
13 877
72
8
12]
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
89 27
22
2
38 19
14 251
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
2
2
2 16
27 239
27
2
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
17 199 132
5
2
2 17
2
7
2
13 100 235 61 483
7 308 16 735
2
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
16
2
2
17
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
2 24
9
4
4
7
29 144
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
8
2
4
28
Next, we use one of the most promising techniques in natural language processing:
word embeddings. In our binary classifier from Recipe 20.3, we one-hot encoded the
observations and used that as inputs to the neural network. However, this time we
will represent each word as a vector in a multidimensional space, and allow the dis‐
tance between two vectors to represent the similarity between words. In Keras we can
do this by adding an Embedding layer. For each value sent to the Embedding layer, it
will output a vector representing that word. The following layer is our LSTM layer
with 128 units, which allows for information from earlier inputs to be used in futures,
directly addressing the sequential nature of the data. Finally, because this is a binary
classification problem (each review is positive or negative), we add a fully connected
output layer with one unit and a sigmoid activation function.
It is worth noting that LSTMs are a very broad topic and the focus of much research.
This recipe, while hopefully useful, is far from the last word on the implementation of
LSTMs.
20.17 Classifying Text
|
335
CHAPTER 21
Saving and Loading Trained Models
21.0 Introduction
In the last 20 chapters and around 200 recipes, we have covered how to take raw data
and use machine learning to create well-performing predictive models. However, for
all our work to be worthwhile we eventually need to do something with our model,
such as integrating it with an existing software application. To accomplish this goal,
we need to be able to both save our models after training and load them when they
are needed by an application. That is the focus of our final chapter together.
21.1 Saving and Loading a scikit-learn Model
Problem
You have a trained scikit-learn model and want to save it and load it elsewhere.
Solution
Save the model as a pickle file:
# Load libraries
from sklearn.ensemble import RandomForestClassifier
from sklearn import datasets
from sklearn.externals import joblib
# Load data
iris = datasets.load_iris()
features = iris.data
target = iris.target
# Create decision tree classifer object
classifer = RandomForestClassifier()
337
# Train model
model = classifer.fit(features, target)
# Save model as pickle file
joblib.dump(model, "model.pkl")
['model.pkl']
Once the model is saved we can use scikit-learn in our destination application (e.g.,
web application) to load the model:
# Load model from file
classifer = joblib.load("model.pkl")
And use it make predictions:
# Create new observation
new_observation = [[ 5.2, 3.2, 1.1, 0.1]]
# Predict observation's class
classifer.predict(new_observation)
array([0])
Discussion
The first step in using a model in production is to save that model as a file that can be
loaded by another application or workflow. We can accomplish this by saving the
model as a pickle file, a Python-specific data format. Specifically, to save the model we
use joblib, which is a library extending pickle for cases when we have large NumPy
arrays—a common occurrence for trained models in scikit-learn.
When saving scikit-learn models, be aware that saved models might not be compati‐
ble between versions of scikit-learn; therefore, it can be helpful to include the version
of scikit-learn used in the model in the filename:
# Import library
import sklearn
# Get scikit-learn version
scikit_version = joblib.__version__
# Save model as pickle file
joblib.dump(model, "model_{version}.pkl".format(version=scikit_version))
['model_0.11.pkl']
338
|
Chapter 21: Saving and Loading Trained Models
21.2 Saving and Loading a Keras Model
Problem
You have a trained Keras model and want to save it and load it elsewhere.
Solution
Save the model as HDF5:
# Load libraries
import numpy as np
from keras.datasets import imdb
from keras.preprocessing.text import Tokenizer
from keras import models
from keras import layers
from keras.models import load_model
# Set random seed
np.random.seed(0)
# Set the number of features we want
number_of_features = 1000
# Load data and target vector from movie review data
(train_data, train_target), (test_data, test_target) = imdb.load_data(
num_words=number_of_features)
# Convert movie review data to a one-hot encoded feature matrix
tokenizer = Tokenizer(num_words=number_of_features)
train_features = tokenizer.sequences_to_matrix(train_data, mode="binary")
test_features = tokenizer.sequences_to_matrix(test_data, mode="binary")
# Start neural network
network = models.Sequential()
# Add fully connected layer with a ReLU activation function
network.add(layers.Dense(units=16,
activation="relu",
input_shape=(number_of_features,)))
# Add fully connected layer with a sigmoid activation function
network.add(layers.Dense(units=1, activation="sigmoid"))
# Compile neural network
network.compile(loss="binary_crossentropy", # Cross-entropy
optimizer="rmsprop", # Root Mean Square Propagation
metrics=["accuracy"]) # Accuracy performance metric
# Train neural network
history = network.fit(train_features, # Features
21.2 Saving and Loading a Keras Model
|
339
train_target, # Target vector
epochs=3, # Number of epochs
verbose=0, # No output
batch_size=100, # Number of observations per batch
validation_data=(test_features, test_target)) # Test data
# Save neural network
network.save("model.h5")
Using TensorFlow backend.
We can then load the model either in another application or for additional training:
# Load neural network
network = load_model("model.h5")
Discussion
Unlike scikit-learn, Keras does not recommend you save models using pickle.
Instead, models are saved as an HDF5 file. The HDF5 file contains everything you
need to not only load the model to make predictions (i.e., architecture and trained
parameters), but also to restart training (i.e., loss and optimizer settings and the cur‐
rent state).
340
|
Chapter 21: Saving and Loading Trained Models
Index
A
accuracy, 186-188
AdaBoostClassifier, 247-249
AdaBoostRegressor, 247-249
adaptiveThreshold, 137-140
affinity, 295
Agglomerative clustering, 294-295
algorithm, 255
algorithms for faster model selection, 219-220
ANOVA F-value statistic, 174, 176
append, 56
apply, 54, 55
area under the ROC curve (AUCROC), 192
arrays
descriptive statistics about, 8
max and min methods, 7
one-dimensional, 1
reshaping, 9-10
slicing, 126-128
two-dimensional, 2
augment_images, 332
average, 8, 42
axis, 8, 55, 56
B
back-filling, 120
backpropagation, 298
Bag of Words model, 104-106
balanced, 245, 265
bandwidth, 292
baseline classification model, 184-186
baseline regression model, 183-184
base_estimator, 248
batch_size, 291, 304
Bayes' theorem, 279-286
(see also naive Bayes classifiers)
Beautiful Soup, 97
Bernoulli naive Bayes, 283-284
best_estimator_, 215
best_params, 324
BigramTagger, 103
Binarizer, 73
binarizing images, 137-140
binary classifier thresholds, 189-192
binary classifiers, 259-260, 301, 303-305
binary feature variance thresholding, 171
bins, 74
blurring images, 128-130
boolean conditions, 112
bootstrap, 240, 241
Brown Corpus text tagger, 103
C
C, 210
calibrating predicted probabilities, 284-286
callbacks, 316, 319
Canny edge detector, 144-146
categorical data, 81-94
dictionary encoding, 86-88
imbalanced classes, 90-94
imputing missing values, 88-90
nominal features encoding, 82-84
nominal versus ordinal, 81
ordinal features encoding, 84-86
chi-squared statistics, 174-176
class separability maximization, 162-164
classes, 102
classification_report, 203
341
classifier prediction evaluation, 186-188
classifier__, 215
cleaning and parsing HTML, 97
cleaning text, 95-96
clustering, 74-76, 287-295
Agglomerative, 294-295
cluster_centers_, 290
DBSCAN, 292-293
evaluating clustering models, 198-199
k-means, 75, 288
meanshift, 291
mini-batch k-means, 290
color isolation, 135-137
compile, 302
concat, 55
confusion matrix, 194-196
contamination, 70
contrast in images, 133-135
convolutional neural networks (ConvNets),
327-332
corner detection, 146-150
correlation matrix, 172
count, 42
CountVectorizer, 104-106
criterion, 236
cropping images, 126-128
cross-validation
cross_val_score, 221
cross-validation (CV), 178, 179, 193
cross_val_score, 182, 186-188, 322
GridSearchCV, 210-211
k-fold cross-validation, 321-322
nested, 220-222
out-of-bag (OOB) observations as alterna‐
tive to, 250
CSR matrices, 4
CSV files, 27
custom evaluation metric, 199-201
cv, 182
D
data
categorical (see categorical data)
defined, xiii
missing values, 76-79
data loading, 23-31
CSV files, 27
Excel file, 28
JSON files, 29
342
|
Index
sample datasets, 23-24
simulated data, 25-27
from SQL database, 30
data wrangling, 33-60
(see also DataFrames)
DataFrames, 33-60
applying function over all elements, 54
applying function to groups, 55
concatenating, 55-57
conditional statements, 38
creating, 34
deleting columns, 46-47
deleting rows, 47
describing data in, 35-37
descriptive statistics and, 42-43
dropping duplicate rows, 48-49
grouping rows, 50-53
index values, 38
looping over a column, 53
merging, 57-60
missing values selection, 44-46
navigating, 37-38
renaming columns, 41-42
replacing values, 39-41
unique values in, 43
dates and times (datetimes) (see time series
data)
DBSCAN clustering, 292-293
decision trees
controlling tree size, 246-247
DecisionTreeClassifier, 233-234
DecisionTreeRegressor, 235-236
model visualization, 236-238
deep learning, 298
describe, 36, 36
determinants, 12
diagonal, 13
dictionaries of features, 86-88
dictionary of candidate learning algorithms,
214-215
DictVectorize, 86-88
dimensionality reduction, 157-167
(see also feature extraction; feature selec‐
tion)
discreditization, 73
distance, 255
distance metrics, 253
document frequency, 107
DOT format, 236-238
dot products, 16
downsampling, 92, 94
drop/drop_duplicates, 46, 48-49
dropout, 317
DummyClassifier, 184-186
dummying, 84
DummyRegressor, 183-184
E
early stopping, 315-317
edge detection, 144-146
Eigenvalues/Eigenvectors, 15
elements, applying operations to, 6
epochs, 304
eps, 293
Euclidean distance, 253
evaluating models (see model evaluation)
Excel files, 28
explained_variance_ratio_, 163
F
false positive rate (FPR), 191
feature creation, 150-152
feature extraction, 157-167
linear discriminant analysis (LDA), 162-164
linearly inseparable data, 160-162
non-negative matrix factorization (NMF),
165-166
principal component analysis (PCA),
158-162
Truncated Singular Value Decomposition
(TSVD), 166-167
feature reduction, 231-232
feature selection, 157, 169-178
binary feature variance thresholding, 171
highly correlated features, 172-174
irrelevant features removal, 174-176
methods for, 169
recursive feature elimination (RFE),
176-178
variance thresholding (VT), 170-172
features_pca__n_components, 217
FeatureUnion, 216
feature_importances_, 241-243
feedforward neural networks, 297, 300
filepath, 320
filter2D, 131
fit, xiii, 62
fit_generator, 332
fit_transform, 62
flatten, 11
flow_from_directory, 332
for loops, 54
forests (see random forests; tree-based models)
forward propagation, 298
forward-filling, 120
FunctionTransformer, 68, 69
G
Gaussian naive Bayes, 280-282
get_feature_names, 87
goodFeaturesToTrack, 148
GrabCut, 140-144
gridsearch.fit, 221
GridSearchCV, 210-211, 216, 221, 256-257,
322-324
groupby, 50-51, 55
H
Harris corner detector, 146-150
HDF5 model, 339-340
head, 35, 36
hierarchical merging, 294-295
highly correlated features, 172-174
histogram equalization, 133-135
histograms, 153-156
hold-out, 180
HTML, parsing and cleaning, 97
hyperparameters, xiii
selecting, 322-324
tuning/optimizating (see model selection)
value effects, 205-207
hyperplanes, 267, 276-277
I
iloc, 37, 38
image augmentation, 331
image classification, 121-156
background removal, 140-144
binarizing, 137-140
blurring, 128-130
color histograms as features, 153-156
contrast, 133-135
corner detection, 146-150
cropping, 126-128
edge detection, 144-146
feature creation, 150-152
Index
|
343
isolating colors, 135-137
loading, 122-124
mean color as a feature, 152
resizing, 125
saving, 124
sharpening, 131-132
ImageDataGenerator, 332
imbalanced classes, 245, 277
imputation, 78, 88-90
Imputer, 78
imwrite, 124
index slicing, 112
interaction features, 66
interactive effects, 225-227
interpolation, 119
IQR, 70
irrelevant classification features removal,
174-176
isnull, 44
J
JSON files, 29
K
k-fold cross-validation (KFCV), 181-182,
321-322
k-means clustering, 75, 288
k-nearest neighbors (KNN), 78-79, 89-90,
251-258
creating a KNN classifier, 254-255
finding best k value, 256-257
locating, 251-254
radius-based (RNN) classifier, 257-258
Keras, 298
convolutional neural networks with,
327-331
Dropout, 317-319
EarlyStopping, 315-317
fit method of training, 304
GridSearchCV, 322-324
input_shape, 302
model saving and loading, 339-340
ModelCheckpoint, 319-320
model_to_dot, 325-326
plot_model, 325-326
predict, 309-310
scikit-learn wrapper, 321-322
Sequential, 300-302
softmax activation, 305-307
344
|
Index
training for regression, 307-308
weight regularization, 313-315
KerasClassifier, 322
kernelPCA, 160-162
kernels, 128-130, 131
for nonlinear decision boundaries, 270-274
for non-linear dimensionality reduction,
160-162
KNeighborsClassifier, 254-255
kneighbors_graph, 253
kurtosis, 43
L
L1/L2 norm, 66
LabelBinarizer, 82
lagged time features, 116
lasso regression, 229-232
learning algorithms, xiii
learning curves, 201-203
learning_rate, 248
limit_direction, 120
linear discriminant analysis, 162-164
linear regression, 223-232
feature reduction, 231-232
fitting a line, 223-225
interactive effects, 225-227
nonlinear relationships in, 227-229
variance reduction with regularization,
229-231
LinearDiscriminantAnalysis, 164
LinearSVC, 267-270
linkage, 295
loading images, 122-124
loc, 37, 38
logistic regression, 259-265
binary classifiers, 259-260
and large data sets, 263
multinomial (MLR), 261-262
variance reduction with regularization,
262-263
LogisticRegressionCV, 219-220, 263
long short-term memory (LSTM) recurrent
neural network, 333-335
loops, 54
loss/loss functions, xiii, 298, 301-302
M
make_blobs, 26-27
make_circles, 160-162
make_classification, 25-27
make_regression, 25-26
make_scorer, 200-201
Manhattan distance, 253
Matplotlib, 122, 310-313
matrices
adding/subtracting, 17
calculating trace, 14
compressed sparse row (CSR), 4
confusion matrix, 194-196
creating, 2-4
describing, 6
determinants, 12
diagonal elements, 13
factorization, 165-166
finding Eigenvalues/Eigenvectors, 15
flattening, 12
inverting, 19
multiplying, 18
rank, 12
selecting elements in, 4
m1sei, 5
sparse, 3, 105
max pooling, 330
maximum/minimum values, 7, 42
max_depth, 247
max_features, 239, 241
max_output_value, 139
mean, 42
mean color as feature, 152
mean squared error (MSE), 196-198
meanshift clustering, 291
median, 43
merge, 57-60
metric, 253, 255, 293
mini-batch k-means clustering, 290
Minkowski distance, 253
MinMaxScaler, 61-62
min_impurity_split, 247
min_samples, 293
Missing At Random (MAR), 77
Missing Completely At Random (MCAR), 77
missing data, 76-79
missing data in time series, 118-120
Missing Not At Random (MNAR), 77
mode, 43
model evaluation, 179-207
baseline classification model, 184-186
baseline regression model, 183-184
binary classifier prediction evaluation,
186-188
binary classifier thresholds, 189-192
classification report, 203
classifier performance visualization,
194-196
clustering models, 198-199
cross-validation (CV), 179
custom evaluation metric, 199-201
hyperparameter value effects, 205-207
multiclass classifier predictions, 192-194
regression models, 196-198
training set size visualization, 201-203
model selection, 209-222
algorithm-specific methods for speed,
219-220
exhaustive search, 210-211
from multiple learning algorithms, 214-215
parallelization for speed, 217-218
post-selection performance evaluation,
220-222
preprocessing during, 215-217
randomized search, 212-213
model.pkl, 338, 338
ModelCheckpoint, 317, 319-320
models, xiii
model_to_dot, 325-326
moving time windows, 117-118
multiclass classifier predictions, 192-194
multiclass classifiers, 301, 305-307
multinomial logistic regression (MLR), 261-262
multinomial naive Bayes, 282-283
N
n-grams, 105
naive Bayes classifiers, 279-286
Bernoulli, 283-284
calibrating predicted probabilities, 284-286
Gaussian, 280-282
multinomial, 282-283
Natural Language Toolkit (NLTK), 99
PorterStemmer, 100
pre-trained parts of speech tagger, 101-103
stopwords, 99
word_tokenize, 98
neg_mean_squared_error, 197
nested cross-validation (CV), 220-222
neural networks, 297-335
binary classification, 303-305
Index
|
345
convolutional, 327-332
deep, 298
designing, 300
dropout, 317
early stopping, 315-317
feedforward, 297, 300
hyperparameter selection, 322-324
image augmentation, 331
image classification, 327-331
k-fold cross-validation (KFCV), 321-322
making predictions, 309-310
multiclass classifiers, 305-307
overfitting, reducing, 313-319
preprocessing data for, 298-299
recurrent, 333-335
regression training, 307-308
saving model training process, 319-320
text data classification, 333-335
training history visualization, 310-313
visualizing, 325-326
weight regularization, 313-315
nominal categorical data, 81
non-negative matrix factorization (NMF),
165-166
nonlinear decision boundaries, 270-274
Normalizer, 64-66
normalizing observations, 64-66
notnull, 44
numerical data
clustering observations, 74-76
discreditization, 73
imputing missing values, 78-79
observations with missing values, 76-77
observations, normalizing, 64-66
outliers, detecting, 69-71
outliers, handling, 71-72
polynomial and interaction features, 66-68
rescaling, 61-62
standardizing, 63-64
transforming features, 68
NumPy
add/subtract, 17
creating matrices in, 2-4
creating vectors in, 1
for deleting missing values, 76
describing matrices in, 6
det, 12
diagonal, 13
dot, 16, 18
346
|
Index
flatten, 11, 150
inv, 19
linalg.eig, 15
matrix_rank, 12
max and min, 7
mean, var, and std, 8
NaN, 45, 72
offset, 14
random, 20
reshape, 9-10, 12
selecting elements in, 4-5
trace, 14
transpose, 10
vectorize function, 6
n_clusters, 289, 295
n_components, 159, 164, 166, 167
n_estimators, 240, 241, 248
n_iter, 213
n_jobs, 218, 255
n_jobs=-1, 182
n_support_, 277
O
observations, xiii
clustering, 74-76
deleting those with missing values, 76-77
offset, 14
one-hot encoding, 82-84
one-vs-rest logistic regression (OVR), 261-262
Open Source Computer Vision Library
(OpenCV), 121
Canny edge detector, 144-146
cornerHarris, 146-150
equalizeHist, 133-135
imread, 122
imwrite, 124
optimizers, 302
ordinal categorical data, 81, 84-86
out-of-bag (OOB) observations, 249
outliers
detecting, 69-71
handling, 71
outlier_label, 258
overfitting, 313-319
P
pad_sequences, 334
pandas
apply, 68, 69
create_engine, 31
DataFrame object (see DataFrames)
descriptive statistics, 42-43
for deleting missing values, 76
json_normalize, 30
read_csv, 28
read_excel, 28
read_json, 29
read_sql_query, 30, 31
rolling, 117
Series.dt, 113, 115
shift, 116
TimeDelta, 114
to_datetime, 109
transformation in, 68
tz_localize, 111
weekday_name, 115
parallelization, 217-218
parameters, xiii
parsing and cleaning HTML, 97
Penn Treebank tags, 102
performance, xiii
performance boosting, 247-249
performance evaluation, 220-222
pickle model, 337-338
Platt scaling, 275
plot_model, 325-326
polynomial regression, 227-229
PolynomialFeatures, 66-68, 226-229
pooling layers, 330
PorterStemmer, 100
pos_tag, 103
precision, 187
predicted probabilities, 274-276, 284-286
predictions, 309-310
preprocess, 217
preprocessing steps in model selection, 215-217
principal component analysis (PCA), 158-162
punctuation, removing, 98
R
Radius-based (RNN) classifier, 257-258
random forests, 233
(see also tree-based models)
comparing feature importance, 241-243
out-of-bag (OOB) observations, 249
RandomForestClassifier, 91, 203, 207,
238-240
RandomForestRegressor, 240-241
selecting important features, 243-244
random variables, 20
RandomizedSearchCV, 212-213
rank, 12
recall, 187
Receiving Operating Characteristic (ROC)
curve, 189-192
rectified linear unit (RELU), 301
recurrent neural network, 333-335
recursive feature elimination (RFE), 176-178
regression function, 301
regression model evaluation, 196-198
regression training, 307-308
regularization, 229-231, 262-263, 318
regularization penalty, 210
rename, 41
resample, 51-53
rescaling, 61-62
reshape, 9-10, 12
resize, 125
resizing images, 125
RFECV, 176-178
RGB versus GBR, 124
ridge regression, 229-231
RobustScaler, 64
rolling time windows, 117-118
S
saving images, 124
score, 184
scoring, 182
search, exhaustive, 210-211
SelectFromModel, 243-244
Series.dt, 113
shape, 36, 304
sharpening images, 131-132
Shi-Tomasi corner detector, 148
show_shapes, 326
shrinkage penalty (see regularization)
silhouette coefficient, 199
silhouette_score, 199, 199
skewness, 43
slicing arrays, 126-128
softmax activation, 305-307
sparse data feature reduction, 166-167
sparse matrices, 105
SQL queries, 30
sqrt, 240
standard deviation, 8, 43, 63
Index
|
347
standard error of the mean, 43
standardization, 63
standardizer, 182
StandardScaler, 63, 298-299
stochastic average gradient (SAG) solver, 263
stopwords, 99
strategy, 184, 185
stratified, 185
stratified k-fold, 181
strings, converting to dates, 109
sum, 42
support vector classifiers (SVCs), 267-270
support vector machines, 267-278
identifying support vectors, 276-277
imbalanced classes, 277
kernel functions for nonlinear decision
boundaries, 270-274
predicted probabilities for, 274-276
support vector classifiers (SVCs) and,
267-270
support_vectors_, 277
svd_solver="randomized", 159
T
tail, 36
term frequency, 107
term frequency-inverse document frequency
(tf-idf), 106
text handling, 95-108
cleaning text, 95-96
encoding with Bag of Words model,
104-106
parsing and cleaning HTML, 97
removing punctuation, 98
stemming words, 100
stop words removal, 99
tagging parts of speech, 101-103
weighting word importance, 106-108
word tokenization, 98
TfidfVectorizer, 106
thresholding, 138
time series data, 109-120
calculating difference between dates, 114
converting strings to dates, 109
encoding days of week, 115
lagged features, 116
missing data, 118-120
multiple date feature creation, 113, 114
rolling time windows, 117-118
348
|
Index
selecting dates and times, 112
time zones, 111-112
TimeDelta, 114
tokenization, 98
toy datasets, 24
to_datetime, 109
trace, 14
train, xiii
training set size effects, 201-203
transform, 62
transforming features, 68
translate, 98
transposing, 10
tree-based models, 233-250
boosting, 247-249
controlling tree size, 246-247
decision tree classifier, 233-234
decision tree regressor, 235-236
imbalanced classes, 245
out-of-bag (OOB) observations, 249
random forest classifier, 238-240
random forest feature importance, 241-243
random forest feature selection, 243-244
random forest regressor, 240-241
visualizing, 236-238
TrigramTagger, 103
true positive rate (TPR), 191
Truncated Singular Value Decomposition
(TSVD), 166-167
tz_localize, 111
U
uniform, 186
UnigramTagger, 103
unique, 43
upsampling, 92
V
validation, 180
validation curve, 205-207
validation_curve, 207
validation_data, 305
validation_split, 305
value_counts, 43
variance, 8, 43
variance reduction, 262-263
variance thresholding, 170-172
vectorize, 6
vectors
calculating dot products, 16
creating, 1
selecting elements in, 4-5
verbose, 211, 304
visualization of classifier performance, 194-196
weight regularization, 313-315
weights, 255
whiten=True, 159
word embeddings, 335
word tokenization, 98
W
weekday_name, 115
Index
|
349
About the Author
Chris Albon is a data scientist and political scientist with a decade of experience
applying statistical learning, artificial intelligence, and software engineering to politi‐
cal, social, and humanitarian efforts—from election monitoring to disaster relief.
Currently, Chris is the Chief Data Scientist at BRCK, a Kenyan startup building a rug‐
ged network for frontier market internet users.
Colophon
The animal on the cover of Machine Learning with Python Cookbook is the Narina
trogon (Apaloderma narina), which is named for the mistress of French ornithologist
François Levaillant, who derived the name from a Khoikhoi word for “flower”, as his
mistress’s name was difficult to pronounce. The Narina trogon is largely found in
Africa, inhabiting both low and highlands, and tropical and temperate climates, usu‐
ally nesting in the hollows of trees. Its diverse range of habitats makes it a species of
least conservation concern.
The Narina trogon eats mostly insects and small invertebrates as well as small rodents
and reptiles. Males, which are more brightly colored, give off a grating, low, repeated
hoot to defend territory and attract mates. Both sexes have green upper plumage and
metallic blue-green tail feathers. Female faces and chest plumages are brown, while
males have bright red undersides. Immature birds have similar coloring to females
with distinct white tips to their inner wings.
Many of the animals on O’Reilly covers are endangered; all of them are important to
the world. To learn more about how you can help, go to animals.oreilly.com.
The cover image is from Wood’s Animate Creation. The cover fonts are URW Type‐
writer and Guardian Sans. The text font is Adobe Minion Pro; the heading font is
Adobe Myriad Condensed; and the code font is Dalton Maag’s Ubuntu Mono.
Download